Datasets:

babytreecc

/

ICLR2025review

like 7

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 2 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "My main questions are about Section 6, as stated in the weaknesses part." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. Clear statement of setup and theoretical results.\n\n2. Detailed proof with several illustrations of proof steps via pictures.\n\n3. Previous results are clearly mentioned with detailed references.\n\n4. The results in this paper extend previous findings under convexity to weaker conditions, which is an important improvement." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper studies the stability of SGLD, which implies generalization and differential privacy guarantees of SGLD. Instead of assuming strong convexity of the loss function, the authors demonstrate that stability results still hold under the dissipativity assumption. Technically, their result is established via verify the uniform LSI of SGLD outputs. Beyond the dissipativity assumption, they also establish a stability result via utilizing the regularizing properties of Gaussian convolution." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The writing in some parts is confusing, making it difficult to clearly understand the contribution in Section 6:\n\n1. In line 199, the authors state, \"we assume in the following that the Gibbs distribution with density proportional to exp(-Fn) satisfies the LS,\" but in Assumption 19, the authors seem to state this LSI assumption again. Is there any difference?\n\n2. In Section 6.1, the authors seem to claim two important preliminary results in Lemma 16 and 17 but don't explain how they affect establishing the main result.\n\n3. It seems that the results in Section 6 are established without verifying the uniform LSI. If so,I am wondering if the analysis template in Section 4 is only applied in Section 5 and whether it should be merged with Section 5. Moreover what is the main proof framework for establishing results in Section 6?\n\n\nOther minor writing problems\n\n1. In line 90, should it be \"the bound does not decay to zero\"?\n\n2. In lines 439, 452, 874, \"given in Theorem 12.\"\n\n3. In line 504, \"given in equation 8 and equation 9.\"" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "I'm not sure if I understood the results in Section 6, which seems to be adopting an entirely different style of analysis as compared to section 5, which helps in lifting the LSI assumption on the entire sequence of intermediate distributions to LSI on the Gibbs distribution corresponding to the loss function. It would help if this approach is explained more thoroughly to see the idea in there a bit more clearly.\n\nI'm open to increasing my score, especially if Section 6 has some good ideas that I might have missed." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "- The paper is well-written and explains its ideas in a well-paced manner, introducing a bit of background, assumptions, and supporting theorems at convenient locations for the reader to follow.\n- The paper presents an example where non-convex loss can provide generalization guarantees that are non-vacuous in number of iterations.\n- The paper simplifies the expansive-contractive decomposition of SGLD steps used in related works for bounding information divergence." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper explores KL and Rényi divergence stability of Stochastic Gradient Langevin Dynamics (SGLD) algorithm. The main characteristic of the presented stability bounds is that they do not become vacuous with the number of iteration of SGLD, which is achieved by assuming log-Sobolev type isoperimetric inequality being satisfied, either throughout the stochastic process, or just by the steady-state Gibbs distribution that SGLD asymptotically approximates. Such isoperimetric properties have also been recently shown to provide rapid convergence in informational divergence as well as convergent DP properties. In a similar vein, the paper derives non-asymptotic and convergent generalization bounds for SGLD as well as bounds on Rényi DP under isoperimetric assumptions. Moreover, the paper shows that the isoperimetric assumption is satisfied under settings considerably milder than strongly-convex losses, such as under dissipative and smooth losses." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Non-analytical issues:\n\n- There is no comparison of the presented bounds with existing results in literature. The generalization bounds presented should be compared to both information-theoretic and non-information theoretic bounds under similar sets of assumptions.\n\n- Contrasting with lower bounds on stability is needed to assess gaps in the tightness of the analysis presented. If such an analysis proves difficult, a well-designed experimental evaluation to compare the generalization bounds with the actual generalization behaviour under the stated assumptions should have been included.\n\nTechnical comments:\n\n- Firstly, information-theoretic generalization bounds inspired by Xu and Raginsky seem to have an O(1/sqrt(n)) dependence on the dataset size, even in cases where other generalization approaches give a better O(1/n) bounds [1]. Since the bounds presented in this paper show dependence in the dataset size n only through Lemma 2 (by Xu and Raginsky), I believe the paper's generalization guarantees might have suboptimal dependence on n under the assumptions made.\n\n- Lemma 3 for conversion of Rényi DP to $(\\epsilon, \\delta)$-DP isn't the best known bound. [Theorem 21, 2] gives a strict improvement which is the best known conversion in my knowledge.\n\n- While Theorem 7 neatly presents the change in Rényi divergence under LSI after a single SGLD step, I believe this inequality might be loose, specially in the dependence on the order $q$ of Rényi divergence. That's because the paper slightly modifies the expansion-contraction template used in other prior works for simplicity. In [3] the expansion-contraction step seems to occur simultaneously, which yield a PDE that is better able to quantify the change in Rényi divergence when integrated over a single step.\n\n- In Section 5.1, the constant of LSI under convexity is dimension independent. But on relaxing strong convexity to dissipativity, the LSI constant has an exponential dependence O(e^d) on the dimension size. The paper further claims in line 418 that this dependence on dimension can't be improved without additional assumptions. To me, this seems like a major hurdle that greatly limits the applicability of the generalization bounds presented (both Corollary 14.1 and 15.1) as plugging in the $C_{LSI}$ constant of Theorem 12 gives an $KL(X_t\\Vert X'_t) = O(e^d)$ dependence on dimension $d$. \n\n\n[1] Haghifam, Mahdi, et al. \"Limitations of information-theoretic generalization bounds for gradient descent methods in stochastic convex optimization.\" International Conference on Algorithmic Learning Theory. PMLR, 2023.\n\n[2] Balle, Borja, et al. \"Hypothesis testing interpretations and renyi differential privacy.\" International Conference on Artificial Intelligence and Statistics. PMLR, 2020.\n\n[3] Chourasia, Rishav, Jiayuan Ye, and Reza Shokri. \"Differential privacy dynamics of langevin diffusion and noisy gradient descent.\" Advances in Neural Information Processing Systems 34 (2021): 14771-14781.\n\nCrafting examples of loss functions satisfying the assumptions made and computing a lower bound on how the KL or Rényi divergence changes with iterations seems doable." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "(1) In the second paragraph under contributions, \"all the iterates of verify a uniform log-Sobolev inequality''. Should it be \"all the iterates verify a uniform log-Sobolev inequality''?\n\n(2) In Lemma 2, is the constant $c$ the same one from Assumption 1? If so, you should mention in the statement of Lemma 2 that you are assuming Assumption 1. For a related question, for each lemma and theorem, it would be really nice if the author(s) can make it more transparent which assumptions are used, especially because the paper contains quite many theoretical results in different settings which require different assumptions.\n\n(3) It would be nice if the author(s) can add some intuitive explanations about the half-step technique in the analysis. For example, when you split the Gaussian noise $N_{k+1}$ into $N_{k+1}^{(1)}$ and $N_{k+1}^{(2)}$, why the former becomes expansive, whereas the latter becomes contractive.\n\n(4) Assumption 14 seems to be a bit strange. If it is pseudo-Lipschitz, shouldn't it be small\nwhen $z$ and $z'$ are close to each other but I do not see $z$ and $z'$ appearing on the right hand side." }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "(1) The paper is well written, and it is a very solid theoretical paper. \n\n(2) The bounds are uniform-in-time, obtained under Renyi divergence under dissipativity condition and KL stability without dissipativity.\n\n(3) A key ingredient in the proof is to show that under dissipativity, all the iterates verify a uniform LSI, which was previously shown only in the strongly-convex setting. This by-product resolves an open question in the literature." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper studies generalization of stochastic gradient Langevin dynamics (SGLD) via information theoretic bound. The author(s) obtained Renyi stability by assuming the iterates verify the log-Sobolev inequality (LSI). The author(s) further showed that the LSI indeed is satisfied under some dissipativity condition. Further results are obtained when dissipativity is not available, in which case KL stability can still be achieved. The bounds are uniform-in-time which are strong. A by-product is that the paper shows that under dissipativity, all the iterates verify a uniform LSI, which was previously shown only in the strongly-convex setting, that resolves an open question in the literature." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "(1) Assumption 15 seems to be a really strong assumption. It would be nice if the author(s) can comment on whether this assumption is needed because of the proof technique or it might be unavoidable.\n\n(2) As the author(s) mentioned in the conclusion section, the dimension dependence is strong. But since the author(s) are working with non-convex setting, this is understandable." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. Isoperimetric vs. Log-Sobolev Inequality:\nThe paper mentions the use of the isoperimetric inequality, but the arguments seem entirely based on the log-Sobolev inequality (LSI). In probability theory, the isoperimetric inequality is usually considered a separate concept. Could this be a typo or an imprecise reference?\n\n2. What is \\Tilde{X}_k' in Theorem 5? Is it a typo, and should it be S_k instead?\n\n3. The $S_k$ is not carefully discussed. In SGLD, when drawing a batch of size $b$, could using a smaller batch size lead to a tighter bound on the KL divergence?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The problem studied is interesting and has practical relevance, with clear motivation provided.\nThe paper is well-structured, with different cases and scenarios analyzed in depth.\nThe results are extensively studied across various settings." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper addresses stochastic optimization, where the goal is to minimize $F(x):=E_{Z\\sim \\nu}f(x;Z)$ for some underlying distribution $\\nu$.\nLet $D$ and $D'$ be two datasets, each consisting of $n$ i.i.d. samples from $\\nu$. Running noisy stochastic gradient descent (SGD) on these two datasets yields sequences $\\{X_k\\}$ and $\\{X_k'\\}$ respectively. It is known that the generalization error scales with the KL divergence between the distributions of $X_k$ and $X_k'$ .\n\nThis paper provides a time-independent upper bound on the KL divergence, even as $k\\to \\infty$. The authors first show that when the log-Sobolev inequality (LSI) holds, an upper bound on the KL divergence can be derived. They further demonstrate that under appropriate conditions, such as dissipativity, the distribution satisfies the LSI, thus leading to the desired bound." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The main proofs rely heavily on existing work, like Theorems 6, 8, and 12.\n2. Some assumptions require further discussion. For instance, Assumption 15 seems restrictive in unbounded domains like $R^d$.\n3. In Theorem 12, the LSI constant scales exponentially with the dimension, which could be problematic for high-dimensional settings." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024generalization,\ntitle={Generalization of noisy {SGD} under isoperimetry},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=0VP3LuzZ8K},\nnote={under review}\n}" }, "abstract": { "value": "We study generalization of iterative noisy gradient schemes on smooth non-convex losses. Formally, we establish time-independent information theoretic generalization bounds for Stochastic Gradient Langevin Dynamics (SGLD) that do not diverge as the iteration count increases. Our bounds are obtained through a stability argument: we analyze the distance between SGLD iterates on two datasets sampled from the same distribution. Our result only requires an isoperimetric inequality to hold, which is merely a restriction on the tails of the loss. We thus relax the assumptions of prior work to establish that the iterates stay within a bounded KL divergence from each other. Under an additional dissipativity assumption, we show that the stronger Renyi divergence also stays bounded by establishing a uniform log-Sobolev constant of the iterates. Without dissipativity, we side step the need for local log-Sobolev inequalities and instead exploit the regularizing properties of Gaussian convolution. These techniques allow us to show that strong convexity is not necessary for finite stability bounds and thus for finite generalization and differential privacy bounds." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "generalization", "langevin", "non-convex", "information theory" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/4b3bc60c70ff45d538fca902b4ebfe4d381f521e.pdf" }, "presentation": null, "primary_area": { "value": "learning theory" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Generalization of noisy SGD under isoperimetry" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Overall, I think this is an interesting paper that focuses on an essential problem in the community, i.g., enabling existing TAP trackers to leverage real videos w/o annotations for training. The idea is somewhat incremental but effectively addresses an essential problem in a simple yet effective way. Thus my current rating is ``accept''. I would like to see more author rebuttal in terms of differences w/ existing pseudo label based approaches as mentioned above." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The paper focuses on an interesting problem in the community, i.e., aiming to explore to train TAP models w/ real videos w/o annotations, since the previous approaches mainly focus on learning w/ synthetic datasets;\n- The proposed RealTracker shows that a simpler architecture and training protocols can outperform SOTA trackers like BootsTAPIR and LocoTrack;\n- The paper is well written and organized;" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a simpler and better point tracking approach by pseudo-learning real videos. Specifically, the proposed approach allows real videos without annotations to be used during training by generating pseudo-labels using off-the-shelf teachers. The proposed approach explores to use real video for training point tracking models w/o annotations. Moreover, the authors also study the scaling law to understand the impact of using more real training videos." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Using pseudo-labels for training trackers is well explored, e.g., for some online learning-based trackers like Dino-Tracker, it uses pre-computed optical flow which provides the pseudo ground truth pixel-level correspondences for online training the tracker. For DinoTracker3, pseudo-labelling is explored. Please illustrate more differences with these trackers for better highlighting the contributions;\n- Are there any specific concerns for choosing a teacher model for pseudo label generation? Does the better teacher model with higher tracking performance commonly lead to better tracking performance? Can a single teacher model well support the tracker learning?\n- In Table 2, the time of the per frame and per tracked point is shown. For the online variant, what’s the overall tracking speed (i.e., fps) given an online testing video?\n- Missing Refs for discussion. For completeness, please include more pseudo-label based tracker training approaches [1,2,3,4] for discussion in the related work.\n\n[1] Progressive Unsupervised Learning for Visual Object Tracking;\n\n[2] Unsupervised Learning of Accurate Siamese Tracking;\n\n[2] DINO-Tracker: Taming DINO for Self-Supervised Point Tracking in a Single Video;\n\n[3] CoTracker3: Simpler and Better Point Tracking by Pseudo-Labelling Real Videos;" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. The terminology \"self-supervised fine-tuning\" is indeed questionable in this context. Using state-of-the-art models from the same domain to generate pseudo-labels for supervision is more aligned with teacher-student learning or pseudo-labeling approaches rather than traditional self-supervised learning, where the supervision signals are typically derived from the data itself without external models.\n\n2. The incorporation of domain adaptation strategies during the fine-tuning process would have significantly enhanced the paper's contribution. This could have included techniques specifically designed to address domain shift and better align feature distributions between source and target domains." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The paper's motivation is well-justified, particularly in its approach to eliminate model redundancies, resulting in a more lightweight yet powerful architecture.\n2. The paper demonstrates effective utilization of unlabeled real-world datasets for training, achieving significant performance improvements through this approach.\n3. The experimental analysis is comprehensive, and the visualization results are particularly impressive in demonstrating the model's capabilities." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "1. The authors address the redundancy in modules of various existing point tracking models and propose RealTracker, a network with simplified architecture that achieves better performance and faster processing speed.\n\n2. The authors leverage existing models to generate pseudo-labels for real video data, enabling effective utilization of unlabeled videos for network fine-tuning, which further enhances performance.\n\n3. The authors analyze the impact of real data scale on the network model's performance, providing insights into the relationship between dataset size and tracking effectiveness." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The methodology appears to be more engineering-oriented rather than theoretically innovative, primarily consisting of combinations and modifications of existing methods. The pseudo-label fine-tuning approach is relatively common. Given this is a **deep learning conference**, the technical contributions seem somewhat limited.\n2. As acknowledged in the limitations section, the model's improvement of performance is heavily dependent on the teacher model's capabilities. This strong reliance on existing methods' performance creates a ceiling effect where the training results are constrained by the teacher model's performance limits, potentially reducing the method's generalizability.\n3. The authors aim to bridge the domain gap using real-world dataset training. However, the paper lacks substantial technical innovation in terms of cross-domain adaptation techniques. The approach merely relies on real-data fine-tuning and teacher model voting effects for enhanced robustness, neither of which represents a significant contribution to the field of domain adaptation. More sophisticated cross-domain strategies or novel technical approaches would have strengthened the paper's contribution in addressing the domain gap problem." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "From Table 2, it appears that the training set does matter in the results, The methods training with Kub+15M performed on average better than the methods trained with Kub, please explain and elaborate. What is the difference?\nWhy does the offline method perform better than the online method, Intuitively I would assume the opposite?\nWhat are the limitations and failure cases?\nTable 6, why does SIFT turn on the best results?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper incrementally builds upon point trackers by producing a better approach that leverages other point trackers to produce supervised training data. In the past other trackers have used synthesized data however this is all based on real data. The results seem to better than other point trackers." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "the approach leverages other point trackers to produce training data for their point tracker. Supposedly less additional training data is required compared to other point trackers. The biggest contribution is that the other trackers use real data and not synthetic data for training. Other approaches in the past have typically used point data for tracking." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "It is not clear on what type of motions were tested, if parallax for motion is required, what about zooming like motions with no parallax, does the method work.\nWhat % of occlusion in terms of coverage of the object and in terms of time occluded were not clearly tested.\nThe limitations and failure cases of the algorithm were not explored." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please follow the weakness. If the issues are addressed, I will improve the rating." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. RealTracker combines valuable ideas from the recent state-of-the-art point trackers and eliminates some unimportant modules.\n2. RealTracker proposes a simple semi-supervised training protocol and achieves better results on several public datasets compared to state-of-the-art trackers.\n3. RealTracker explores the training scaling low via its proposed training protocol." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces the RealTracker, a point tracker that combines several ideas from other related trackers but eliminates some components and simplifies others. RealTracker also designes a semi-supervised training protocol, where real videos are annotated utilizing several off-the-shelf trackers. With this protocol, RealTracker can achieve encouraging results on the Kinetics, RGB-S, and DAVIS datasets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The idea of using trackers to annotate unlabeled datasets, such as [1], is not new.\n2. The authors should use the Kub+15M data to train the CoTracker and TAPTR and verify the proposed method's effectiveness.\n3. To prove the effectiveness of the RealTracker, it is suggested that confidence and visibility be visualized.\n4. More ablation studies are suggested to verify that eliminating some modules in the listed trackers and simplifying some modules is useful, including the computation cost and tracking performance.\n\n[1] Muller M, Bibi A, Giancola S, et al. Trackingnet: A large-scale dataset and benchmark for object tracking in the wild[C]//Proceedings of the European conference on computer vision (ECCV). 2018: 300-317." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose a simple point tracking architecture, RealTracker, along with a pseudo-labelling protocol to improve point tracking models. We outperform BootsTAPIR, the state-of-the-art point tracking model, while using 1000x less real data." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024realtracker,\ntitle={RealTracker: Simpler and Better Point Tracking by Pseudo-Labelling Real Videos},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=0Wl6h2CZeJ},\nnote={under review}\n}" }, "abstract": { "value": "Most state-of-the-art point trackers are trained on synthetic data due to the difficulty of annotating real videos for this task.\nHowever, this can result in suboptimal performance due to the statistical gap between synthetic and real videos. In order to understand these issues better, we introduce RealTracker, comprising a new tracking model and a new semi-supervised training recipe. This allows real videos without annotations to be used during training by generating pseudo-labels using off-the-shelf teachers.\nThe new model eliminates or simplifies components from previous trackers, resulting in a simpler and smaller architecture.\nThis training scheme is much simpler than prior work and achieves better results using 1,000 times less data.\nWe further study the scaling behaviour to understand the impact of using more real unsupervised data in point tracking.\nThe model is available in online and offline variants and reliably tracks visible and occluded points." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Point tracking", "Optical flow", "Motion estimation", "Pseudo labelling" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/f165a50bd890a7ae69459dbc2afd80de7fa6ded6.pdf" }, "presentation": null, "primary_area": { "value": "applications to computer vision, audio, language, and other modalities" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/95ccb3ac9b89574da84f5b498acaf51f3d732567.zip" }, "title": { "value": "RealTracker: Simpler and Better Point Tracking by Pseudo-Labelling Real Videos" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. Typo: Algo 1 line 7, should it be \"OOD\" instead of \"ODD\"?\n2. How do you define \"progressively expanding the randomization range\" for different environment parameters? More specifically, increasing friction by 1% and increasing the agent's mass by 1% may have vastly different impacts on the task difficulties. Could you discuss more on the relative impact of changing each parameter to the environment difficulty?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The overall method is well-motivated and clearly stated.\n2. Weighting samples in the normal dataset and repulsive dataset differently is intuitive and is demonstrated to be effective empirically.\n3. Using a set of critiques and its variances as a measure of environmental uncertainty explore new possibilities from the existing DENN method." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a novel pipeline to detect OOD environment variations and gradually fine-tunning the agent until high confidence safe deployment is possible." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Repulsive locations discussion could be written more formally and thus more concisely. The current version is a bit too dense verbally. I'm also a bit confused by Figure 2, while it's a nice visual, is it something derived from the experiments or is it just a conceptual illustration?\n2. Lack of related works: changing environment parameters to achieve repulsive locations is quite related to the literature on curriculum learning. A good survey to start with is https://arxiv.org/pdf/2003.04960. Also, blindly varying the environmental parameters may lead to unexpected harmful environments dampening the agent training: https://openreview.net/forum?id=hp4yOjhwTs&noteId=vZMeHQbnJK \nI would suggest authors add a subsection for curriculum reinforcement learning in the related work for a more thorough introduction to the problem backgrounds." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "see above." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The problem raised in this paper is important and the experiments are solid." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this paper, a novel RL pipeline, Uncertainty-aware Adaptive RL (UARL), has been proposed to enhance policy generalization across diverse variations of a given environment. UARL frames distribution shifts as OOD issues and integrates a new OOD detection method to quantify uncertainty. This method enables iterative policy fine-tuning, beginning with offline training on a limited state space and gradually expanding to more diverse variations of the same environment through online interactions." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Weakness:\n\n1) The core problem in this paper, i.e., distributional shift, and many important concepts are not discussed in detail. In the introduction, the authors discuss about the theoretical shortcomings of robust RL, safe RL, and the distributional shift in offline2online RL, but they ignore the discussion about the relationships between these concepts. For example, what is the relationship between the robust RL and the distributional shift in offline setting? Furthermore, what is the difference between the distributional shift problems in offline RL and offline2online RL settings? Why the proposed method could successfully solve the problem of distributional shift? Please answer this question from a high-level view.\n\n2) In offline (to online) RL, the uncertainty quantifier is defined clearly as the upper bound of the error produced by the empirical Bellman operator (see [1], Eq.(4.1)). Then we may concern that whether the uncertainty defined in this paper's Eq.(5) has the relationship with the uncertainty quantifier as we have known in [1]? Does it a valid uncertainty quantifier theoretically? The author should discuss about this point.\n\n[1] Jin. et al., Is Pessimism Provably Efficient for Offline RL.\n\n3) In offline RL, there have been many uncertain-aware methods to deal with the distributional shift problem, such as [2] and [3]. In the list two works, they both penalize the OOD actions by the constructed uncertainty quantifiers. So in our view, the method in this work is not beyond the scope of these methods and lack of the sufficient discussion with the advantage over the existing uncertain-aware methods.\n\n[2] Bai. et al., Pessimistic bootstrapping for uncertainty-driven offline reinforcement learning.\n[3] Sun. et al., Model-Bellman Inconsistency for Model-based Offline Reinforcement Learning.\n\n4) the convergence property of the proposed algorithm should be discussed, especially line 8 in Algorithm 2 - what if this condition is never voilated?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. Comparing the computational overhead of your method with that of baseline algorithms would strengthen your work. Could you include this information to provide a clearer understanding of its efficiency?\n\n1. Does this algorithm fall within the scope of Offline Reinforcement Learning? If so, it would be helpful to clarify its placement within the Offline Reinforcement Learning landscape. Enhancing the abstract and introduction to better position the algorithm within this broader context would significantly improve the clarity and impact of your paper.\n\nI am open to raising my score based on these improvements." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- UARL presents a compelling approach to address the challenges of deployment a policy in RL. The progressive expansion of state space via repulsive locations and a balanced replay buffer to manage data distribution shifts are novel and theoretically sound.\n-The usage of an ensemble of diverse critics to perform OOD detection and policy refinement represents a robust methodology that has support from the experimental result" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces Uncertainty-aware Adaptive RL (UARL), an innovative framework to tackle distributional shifts and out-of-distribution (OOD) issues when deploying reinforcement learning (RL) policies in real-world environments. This is accomplished by implementing OOD detection to quantify policy uncertainty and iteratively refine high-uncertainty regions (of the state space), adapting the policy\nfor safe and effective performance deployment. UARL demonstrates several notable advancements,\n- A method for quantifying policy uncertainty using OOD detection.\n- An offline-to-online (O2O) adaptation strategy that balances online and offline data, utilizing a diverse ensemble of critics to better handle distributional shifts.\n- Experiments on MuJoCo continuous control tasks that validate UARL’s effectiveness in terms of performance, robustness, and sample efficiency." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The paper could better highlight its unique contributions compared to existing OOD and ensemble-based offline RL methods. A clearer differentiation of UARL's specific advancements would help underscore its novelty within the landscape of similar approaches.\n\n- The experimental validation, limited to few environments such as the Ant-v4 and HalfCheetah-v4 environments, may not fully capture the method’s effectiveness across a diverse range of tasks. Extending the experiments to include more varied environments would provide a more comprehensive assessment and enhance the generalizability of the results.\n\n- A comparison with recent state-of-the-art methods, such as PBRL[1], RORL[2], would strengthen the empirical evaluation. By benchmarking UARL against PBRL and similar approaches, the paper could provide a more robust validation of its improvements in uncertainty handling and performance stability.\n\n[1] Pessimistic Bootstrapping for Uncertainty-Driven Offline Reinforcement Learning\n[2] RORL: Robust Offline Reinforcement Learning via Conservative Smoothing" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please c.f. the comments above. Besides, I have the following questions,\n\n- In Lines 184-185, you wrote, *disagreement among ensemble models, particularly at the boundary of the training distribution*, what do you exactly mean by *at the boundary of the training distribution*?\n- what are the advantages of the diversity term in Equation 5 compared to other diversity terms? (e.g., the diversity term used in the EDAC paper) The authors ought to justify the advantages of doing so rather than using some other methods.\n- how can the authors tell that the uncertainty measurement provided in this paper is valid? It would be better to compare against some other uncertainty estimation methods and visualize the uncertainty measurement for a better comparison\n- do you have any parameter study on the threshold parameter in Algorithm 2? How it can affect the performance of the agent? Do we need to tune this hyperparameter per task? How can we ensure that the policy is safe when $V_Q \\le {\\rm threshold}$?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "## Pros\n\nThis paper enjoys the following advantages,\n\n- This paper is well-written. The presentation of this paper is very good and of high quality. The figures are very nice and helpful. Some of the illustration figures significantly convey the idea and core design of UARL, e.g., Figure 1, Figure 2\n- This paper is easy to read and easy to follow\n- The authors provide open-source codes in the anonymous website, and I believe that the results reported in this paper are reproducible" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper deals with the distribution shift issue in reinforcement learning (RL). The authors introduce an approach called Uncertainty-aware Adaptive RL (UARL) that enhances policy generalization across diverse variations of a given environment. UARL views distribution shifts as OOD problems and integrates an OOD detection method to quantify uncertainty, i.e., Q-value variance. UARL realizes diversity in critics via the DENN method. The authors claim that UARL enables iterative policy fine-tuning, starting with offline training on a limited state space and progressively expanding to more diverse variations of the same environment through online interactions. The authors demonstrate the effectiveness of UARL through some experiments on continuous control tasks, showing improved performance and sample efficiency compared to existing methods." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "## Cons\n\nDespite the aforementioned advantages, this paper has the following flaws.\n\n- **Offline-to-online RL or off-dynamics RL?** This paper claims that they focus on the offline-to-online setting, while it seems that they actually are dealing with off-dynamics RL [1]. The offline-to-online RL typically refers to training a policy offline from a static dataset and then fine-tuning it with some extra online interactions in the same environment. Instead, the authors conduct experiments by modifying the environmental parameters, which essentially constructs dynamics shifts. The experimental setting resembles that in off-dynamics RL [2,3,4]. It seems unclear to me whether it is suitable to name the paper *offline-to-online* or *off-dynamics*.\n- **Insufficient related work and Limited novelty.** The authors emphasize that the proposed method can enhance the safety and robustness of RL, but it includes too few related works on safe offline/offline-to-online RL and robust offline/offline-to-online RL. Meanwhile, I have doubts about the novelty of this work. The authors progressively increase hyperparameter randomization of the environment (e.g., friction) when the variance between the Q-ensemble is large and terminates until the policy is safe enough to be deployed. Such an idea resembles [5], which progressively adapts its learned policy by modifying the parameters of the environment. Furthermore, I am a bit confused about the benefits of parameter randomization against domain randomization. If the user can adjust the parameters of the environment, then why not directly use domain randomization? I would expect reasonable justifications for the design of the tasks here. Furthermore, the diversity term is not novel, it is borrowed directly from the existing literature. These all together make the contribution of this paper somewhat limited.\n- **Lacking baseline algorithms.** As commented above, this paper claims that it addresses the offline-to-online RL but actually focuses on the off-dynamics RL setting, the authors should include the following baselines,\n - baselines on off-dynamics RL, e.g., [2,3,4]. This is vital to show the effectiveness of the UARL in terms of policy generalization to the target domain\n - RLPD [6], which is a method specially designed for learning with offline data and online interactions with the environment. This baseline is important since it exhibits superior performance given the experimental setting described in this paper (offline data and online interactions). Based on my experience, RLPD can achieve quite strong performance even when there exists dynamics shifts between the offline data and the online environment. Involving this baseline can justify the necessity of the components adopted in UARL (otherwise, one can directly use RLPD for deployment)\n - baselines on safe RL and robust RL. The authors claim that UARL can enhance the safety and robustness of the executed actions, while they do not include any safe RL or robust RL methods for comparison, making it hard to see the rationality and effectiveness of UARL\n - baselines on offline-to-online RL. Unfortunately, this paper also does not include offline-to-online RL methods as valid baseline methods. It is hard to tell the true effectiveness of UARL without these methods, e.g., [7,8,9]\n- (minor) **Lacking theoretical justifications.** There is no theoretical analysis of the UARL. I do not want to blame the authors too much on this point. I understand that this paper may set the focus mainly on the empirical side, but including some theoretical analysis can strengthen this paper.\n- (minor) **Other issues.**\n - in Equation 5, you wrote $R(s,a)$ in the bellman error, while $r$ in the diversity term $\\mathcal{L}_{div}^{RL}$. I think they should be identical, right?\n - the authors do not discuss the limitations of their method in the main text or the appendix. It is important to acknowledge both the advantages and the limitations of the proposed method. \n - the performance improvement of UARL seems limited and incremental on some tasks (e.g., see Figure 3)\n - UARL can still suffer from performance degradation during the fine-tuning phase (e.g., see Figure 5)\n\nGiven the above concerns, I vote for rejection since I believe that this paper needs a significant revision before being accepted for possible publication.\n\n[1] Off-dynamics reinforcement learning: Training for transfer with domain classifiers. ICLR\n\n[2] When to trust your simulator: Dynamics-aware hybrid offline-and-online reinforcement learning. NeurIPS\n\n[3] Cross-domain policy adaptation via value-guided data filtering. NeurIPS\n\n[4] Cross-domain policy adaptation by capturing representation mismatch. ICML\n\n[5] Revolver: Continuous evolutionary models for robot-to-robot policy transfer. ICML\n\n[6] Efficient online reinforcement learning with offline data. ICML\n\n[7] Offline-to-online reinforcement learning via balanced replay and pessimistic q-ensemble. CoRL\n\n[8] Bayesian Design Principles for Offline-to-Online Reinforcement Learning. ICML\n\n[9] Proto: Iterative policy regularized offline-to-online reinforcement learning. Arxiv" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. According to Eq. 5, $R(s, a)$ is not sampled from the dataset $\\mathcal{D}$, but the general Q-function update for offline RL, as in Eq. 1, is to use the sampled $r$. Is this a mistake here?\n2. Is there a performance or computational advantage of UARL over direct processing of $E_\\omega$'s using the Robust RL algorithm or the algorithm with domain randomization techniques? Can this be illustrated experimentally?\n3. Notice that in offline training (1st iteration), EDAC performs much worse than CQL and TD3+BC in many environments, which doesn't seem to match the experimental results in the EDAC article?\n4. The experiments in this paper were all performed in Mujoco, how do we obtain the real-world demonstration dataset $\\mathcal{D}_\\omega$ in the simulation environment like Mujoco?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1.This paper propose a novel OOD detection method and an iterative online policy fine-tuning training framework.\n2.Good experimental results are obtained on Mujoco environments with randomized environmental hyperparameter, verifying the validity of the method.\n3.The writing is good." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a novel approach aimed at solving the OOD problem faced by deploying reinforcement learning strategies in real-world scenarios when the distribution of training environments is shifted. The proposed approach tackles this issue by adopting a new diversity ensemble Q-network approach to OOD detection. Furthermore, the method incorporate an iterative policy fine-tuning method that starts with offline training in the original environment and gradually scales up to more stochastic environments through online interactions. Experimental results show that this approach outperforms the Baseline algorithm in Mujoco environments with randomized environmental hyperparameter and typically requires fewer samples to converge." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. It should be further shown, either theoretically or experimentally, why the diversity term $\\mathcal{L}^{\\text{RL}}_{\\text{div}} $ in Eq. 5 allows the ensemble Q network to learn the value of diversity. Intuitively, minimising $\\text{exp}(-\\Vert Q_i(s, a) - (r + Q_i(s^\\prime, a^\\prime)) \\Vert^2/2\\delta^2)$ will allow the Q network to not converge quickly to a certain value on the repulsive dataset, but it does not guarantee that the ensemble Q network learns diverse values.\n2. Further clarification is needed for how to calculate $V_Q$ in Algorithm 2 and how to calculate critical variance in uncertainty estimation experiments.\n3. The ablation experiments in Appendix B.3 are not detailed enough. The training curves for different parameter combinations should be differentiated to illustrate the algorithm's parameter sensitivity to $\\lambda$ and $\\delta$ during training.\n4. For each $E_i$, the randomized environmental hyperparameter range is determined without a common metric but as a hyperparameter, which may require a lot of time for online tuning for complex scenarios." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024mitigating,\ntitle={Mitigating Distribution Shifts: Uncertainty-Aware Offline-to-Online Reinforcement Learning},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=0WqAnYWi7H},\nnote={under review}\n}" }, "abstract": { "value": "Deploying reinforcement learning (RL) policies in real-world scenarios faces challenges due to distribution shifts from training environments. Past approaches have shown limitations such as poor generalization to out-of-distribution (OOD) variations or requiring extensive retraining on new data. We propose Uncertainty-aware Adaptive RL, UARL, a novel RL pipeline that enhances policy generalization across diverse variations of a given environment. UARL frames distribution shifts as OOD problems and incorporates a new OOD detection method to quantify uncertainty. This approach enables iterative policy fine-tuning, starting with offline training on a limited state space and progressively expanding to more diverse variations of the same environment through online interactions. We demonstrate the effectiveness and robustness of UARL through extensive experiments on continuous control tasks, showing improved performance and sample efficiency as well as reliability in OOD detection compared to existing methods." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Reinforcement learning", "Out-of-distribution detection", "Uncertainty estimation", "Offline RL" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/ca50c0b6885834141e7e8abca8ffa54cab3f4a7c.pdf" }, "presentation": null, "primary_area": { "value": "reinforcement learning" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Mitigating Distribution Shifts: Uncertainty-Aware Offline-to-Online Reinforcement Learning" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "N.A" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "See the above strengths and weaknesses." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The motivation behind this paper is reasonable. By analyzing daily CT and MRI scans in the cardiac and abdominal regions, the authors observed two consistent patterns across certain subjects, leading to the formulation of the Piece-wise Smooth (P-S) Assumption. This assumption leverages physical priors from observed medical image patterns, which is both innovative and plausible, enhancing neural network-based registration tasks by grounding them in realistic assumptions about medical image structures.\n2. The paper provides thorough comparative experiments. The authors test AdaWarp on two registration datasets spanning different modalities and input constraints, which demonstrates robustness and broad applicability." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper leverages prior knowledge observed in medical images to introduce the Piece-wise Smooth (P-S) Assumption as a basis for addressing medical image registration tasks. Specifically, the authors propose AdaWarp, a warping method that utilizes learnable adaptive filtering to register medical scans in line with the P-S assumption. By employing a low-resolution latent representation along with a differentiable bilateral grid, the method achieves a better balance between accuracy and efficiency. Experiments conducted on two registration datasets validate the effectiveness of the proposed approach." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The novelty of this paper does not seem particularly strong. While the method leverages an encoder to extract a latent representation that approximates the deformation field at a low resolution, this approach mainly contributes to the model's efficiency but is not unique. The use of latent feature representations for similar tasks has already become common in the field.\n2. The core of AdaWarp is a differentiable bilateral grid, which naturally incorporates the P-S prior. In implementation, the guidance map aids in processes like splatting, blurring, and slicing. This incremental modification lacks sufficient novelty." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. In Figure 4 and Figure 5, why not include VoxelMorph into comparison? VoxelMorph is the most widely-benchmarked method and has high efficiency with low number of parameters. \n2. There is a recent registration study in CVPR (CorrMLP, Meng et al. 2024), which is based on a totally conflicting motivation against this paper. CorrMLP attempted to capture long-range dependency among full-resolution image details in an efficient approach (using MLPs), while this paper suggests that only low-resolution features are sufficient. So, it’s interesting to compare with the CorrMLP: did the proposed method achieve similar registration accuracy while reducing much computational complexity?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The proposed method bridges the gap in the existing literature focusing on the balance between registration accuracy and computational efficiency, which is capable of enforcing global smoothness while respecting local discontinuities. This paper was well-written with very clear description on methodology." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a learning framework that improves the accuracy-efficiency trade-off in medical image registration by leveraging the piece-wise smooth prior. The proposed method was evaluated on two medical image datasets involving cardiac MRI and abdomen CT images. This method transforms the deformable registration problem into a keypoint detection task and shows potential for segmentation tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The major concern is the research focus of this study, which might not be of sufficient significance in the field of medical image registration. After the introduction of deep learning-based registration methods, e.g., VoxelMorph, existing methods have been very fast in registration, allowing real-time registration using GPUs. Under this situation, only a few studies have specifically focused on improving efficiency, which suggests that this topic might not be the real problem in the community. \n2. Another concern is the generalizability of the P-S assumption. In the study, this assumption was exemplified and evaluated with cardiac MRI and abdomen CT images, where there is no too many complex anatomical structures and local deformations. It’s important to evaluate the proposed method on the well-benchmarked brain MRI registration tasks, in which the P-S assumption may fail." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "I encourage the authors to consider addressing as many of the points highlighted in the weaknesses section as possible. Additionally, while the paper presents an intriguing and novel approach, the clarity and quality of the presentation could benefit from further refinement." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The paper presents AdaWarp, a novel method that integrates the piece-wise smoothness assumption enforcing global smoothness while respecting local discontinuities in a learning framework striking a balance between complexity and accuracy. \n\nMoreover, it demonstrates connections of the adaptive filtering approach with the self attention.\n\nThe experimentation on two challenging registration tasks cardiac and inter-subject abdominal registration demonstrate that AdaWarp outperforms existing methods in accuracy-efficiency and accuracy-smoothness tradeoffs." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a novel method to utilise prior knowledge (piece-wise smooth assumption) to enhance learning based registration striking a balance between computational complexity and accuracy. The performance is evaluated on a cardiac and an abdominal dataset." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Although I believe that the paper attempts to bridge a gap in the literature by incorporating a differentiable bilateral grid within a learning-based registration framework, I would like to point out several weaknesses and raise some questions regarding the experiments.\n\n[A] I would like to invite the authors to elaborate on this statement regarding iterative optimization-based methods: “As a result, these approaches tend to be time-consuming and lack the ability to incorporate contextual information effectively.”\n\n[B] “While high-dimensional filtering can project signals onto arbitrary spaces, we focus on extending by one additional dimension to account for the object boundary.”\n\nWhat is the intuition behind this approach? Is only one additional dimension sufficient? I would like to invite the authors to further elaborate and explain their choice.\n\n[C] The role of the guidance map generator component is unclear. Could the authors please explain why this component is used or needed?\n\n[D] Could the authors clarify whether the same lambda values are used for all methods or if different values are applied? How were these values tuned? Were they also tuned for the baselines?\n\n[E] The proposed method utilizes a diffeomorphic transformation model; however, it is not clear whether the baselines follow the same principle. Could the authors provide a table that explicitly lists the hyperparameters used by each of the baselines along with the transformation model?\n\n[F] The authors chose different baselines for the two datasets, which is puzzling. What is the intuition behind this decision? Is there a reason why this approach was chosen?\n\n[G] The paper presents t-tests for DICE scores but not for other metrics. Is there a reason for this choice? Could the authors extend their t-tests to cover HD95 as well?\n\n[H] “Learning-based methods generally outperform traditional ones in registration accuracy, though with slightly higher SDlogJ.”\n\nDo the authors have any intuition as to why this is the case? Normally, I would expect that iterative optimization methods achieve higher accuracy [1].\n\n[1] Hansen, L. and Heinrich, M.P., 2021. Revisiting iterative highly efficient optimization schemes in medical image registration. In Medical Image Computing and Computer-Assisted Intervention–MICCAI 2021: 24th International Conference, Strasbourg, France, September 27–October 1, 2021, Proceedings, Part IV 24 (pp. 203-212). Springer International Publishing.\n\n[I] For the abdominal dataset, the proposed method uses Convex Adam’s framework with the same segmentation model as a feature extractor. Is there any reason for this choice? Could the model be trained from scratch? Could the authors elaborate on the design choices, including why the architecture differs depending on the dataset?\n\n[J] The code is not available. Are the authors planning to make their code publicly accessible?\n\n[K] Due to the lack of ground truth, registration is evaluated quantitatively with surrogate measures. However, to ensure the registration’s success, it is common practice to inspect the resulting transformed images qualitatively as well. I would like to invite the authors to provide qualitative results for both datasets, as this would substantially strengthen their claims." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. Why were different model structures used for different datasets? What would be the result of using Ada-Res on the Abdomen CT dataset and Ada-Cost on the ACDC dataset? A comparison of these model structures across datasets could help demonstrate their generalizability and clarify why different architectures were chosen for each.\n\n2. In the Abdomen CT dataset, Ada-Cost uses “the same segmentation model for feature extraction.” Was this segmentation model pre-trained? If so, this would make Ada-Cost a semi-supervised registration model. Comparing it with other unsupervised deep learning-based methods would be unfair. Additionally, how exactly was the segmentation model integrated into your model’s structure? Does it replace the \"guidance map generator,\" or is it incorporated elsewhere in the architecture?\n\n3. More references, more baselines and visual evaluations of warped images and warped segmentation masks would be highly valuable. Providing such visual results would help demonstrate the effectiveness of your method in producing sharp boundaries, which cannot be fully illustrated through numerical metrics alone.\n\n4. I would greatly appreciate it if the paper could provide information on the inference and training time of the proposed method. This data would offer more valuable insights into the computational efficiency of the model.\n\n5. Another concern is that the authors selected \"interpretability and explainable AI\" as Primary Area. I’m not sure if this is appropriate since there is no work about interpretability of proposed method." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The integration of the differentiable bilateral grid into the deep learning framework for image registration is highly innovative. It effectively addresses the limitations of traditional smoothness constraints, enabling the model to better handle complex and localized deformations.\n\n2. The paper is well-structured, offering a clear explanation of the proposed methods. It provides detailed descriptions of the differentiable bilateral grid, encoder architecture, and adaptive filtering process. Visual aids, such as Figures 4 and 5, are particularly useful in clarifying complex comparisons.\n\n3. This method presents a promising alternative for resolving the conflict between global smoothness and local deformations, potentially offering improved solutions in certain applications." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents AdaWarp, a novel architecture in medical image registration. The model introduces a piece-wise smooth (P-S) assumption, which exploits the smoothness of intensity variations within anatomical regions while preserving sharp boundaries between organs. This assumption is incorporated into the network through a differentiable bilateral grid, which allows for efficient edge-preserving filtering and reduces computational complexity." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The weaknesses of the paper are primarily in the literature review and experimental sections, which lack sufficient references and baseline comparisons, as well as visual results. These limitations are why I rated the paper as \"fair\" in terms of Presentation and Soundness.\n\n\n1. The paper needs more references in the literature review. The current review only discusses works that do not address the conflict between global smoothness and local deformations. However, this is not the first paper to tackle this problem. Research such as multi-scale registration and patch-wise registration also offers relevant solutions. While these methods may not explicitly incorporate the piece-wise smooth prior, they still manage local deformations while maintaining overall smoothness. The authors should include these references in the background and select baselines from this body of work to show that the proposed method offers a superior solution to the problem.\n\n2. The experiments do not adequately support the claimed advantages of the proposed method. While the paper argues that the model can generate sharp boundaries between organs by incorporating the P-S assumption, it fails to provide visual results to substantiate this key contribution. Relying solely on numerical metrics like Dice, HD95, and SDlogJ does not clearly demonstrate that the model’s output preserves sharp boundaries.\n\n3. The writing in the experiments section is somewhat disorganized. The authors employ significantly different model structures and training strategies, including both unsupervised and semi-supervised approaches (which require further clarification), depending on the dataset. This inconsistency raises concerns about the generalizability of the model across different tasks. Additionally, the experiments lack ablation studies, which are necessary to demonstrate the effectiveness of each component in the proposed methods." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "A novel efficient model for medical image registration using differentiable bilateral grid" }, "_bibtex": { "value": "@inproceedings{\nanonymous2024efficient,\ntitle={Efficient Adaptive Filtering for Deformable Image registration},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=0XT3Lg6S2Q},\nnote={under review}\n}" }, "abstract": { "value": "In medical image registration tasks where targets exhibit piece-wise smooth structures, a well-designed low-resolution data structure can approximate full-resolution deformation fields with minimal accuracy loss. \nThis physical prior, though absent in current literature, can be integrated into neural networks to enhance registration. \nIn this paper, we propose AdaWarp, a novel architecture that leverages the prior for more efficient medical image registration. \nAdaWarp consists of an encoder, guidance map generator, and a differentiable bilateral grid, introducing an edge-preserving low-frequency approximation of the deformation field. \nThis approach reduces computational complexity without sacrificing accuracy.\nExperiments on two registration datasets covering different modalities and input constraints demonstrate that AdaWarp outperforms existing methods in accuracy-efficiency and accuracy-smoothness tradeoffs." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Deformable image registration", "Adaptive filtering", "Bilateral Grid", "Piece-wise Smooth" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/1ee3573d6953e189b620bc635ad895fbe75d3d12.pdf" }, "presentation": null, "primary_area": { "value": "interpretability and explainable AI" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Efficient Adaptive Filtering for Deformable Image registration" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "Intuitively, local patch features do not align with text features and therefore cannot be directly utilized, as studied in [1]. Could the authors provide more discussion or visualizations to illustrate this aspect?\n\n[1] A Closer Look at the Explainability of Contrastive Language-Image Pre-training." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The motivation to use local features is intuitive, as they may contain important details that can enhance model performance.\n\n2. The paper conducts extensive experiments, including comparisons of MP-TPT on two representative benchmarks and ablation studies." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper studies the topic of test-time prompt tuning (TPT), and propose MP-TPT. MP-TPT introduces local patch features as additional visual augmentations, which may be crucial for classification. Additionally, it leverages local visual features to enhance text feature descriptions. Extensive experiments demonstrate the effectiveness of the proposed method." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. **Limited novelty and contribution**: The concept of using local features to enrich image features and image counterparts to enhance text features has already been proposed in [1]. The primary difference in this paper is the implementation of this idea in a test-time prompt tuning scenario. Surprisingly, [1] is not cited or discussed in this paper.\n\n2. **Clarity and organization**: The paper is difficult to follow due to disorganized writing, confusing formulas, and figures and tables that are not self-contained. This impacts its suitability for ICLR acceptance. I list some below:\n\n 1. Line 225: The term \"local visual representation\" is unclear. Is this referring to CLIP patch features? This needs clarification.\n 2. Line 253: Why are classification probabilities referred to as \"cross-modal information\"? Is it simply because they use features from two modalities? What specific information do they contain?\n 3. In Equation (7), The resulting shape is $\\mathbb{R}^{W H \\times d}$. How are $M$ augmented features derived?\n 4. In Equation (7), There are 5 left brackets and 3 right brackets, making the expression difficult to understand.\n 5. In Table 1, how is the inference time calculated? Are the times in seconds? Different datasets with varying classes should have different inference speeds. The table should be self-contained.\n 6. Multiple definitions of $K$: In Line 161, $K$ is defined as the number of classes, while in Line 244 and Equation (6), $K$ is the number of selected regions.\n 7. Undefined terms: $\\boldsymbol{f}^t$ in Equation (7) is not defined. Is it a set or a concatenation of $\\boldsymbol{f}^t_i$?\n 8. The definition of a set in Equation (8) is incorrect. The part “$p\\left(y_k \\mid \\tilde{\\boldsymbol{f}}_i^t\\right)$” after the colon should be removed.\n\n3. **Experimental issues**: \n\n 1. The claims in Line 28 are misleading. MP-TPT did not achieve a 1% improvement over TPT and 4.5 times faster simultaneously. These are achieved by different methods, MP-TPT-L and MP-TPT-S.\n\n 2. Some highly relevant works, such as [2] and [3], are missing from Tables 1 and 2. The performance of MP-TPT is significantly lower compared to these methods. More discussion is needed.\n\n | Methods | Cross-dataset | Domain Generalization |\n | --------------- | ------------- | --------------------- |\n | PromptAlign [2] | 66.92 | 63.55 |\n | TDA [3] | 67.53 | 63.89 |\n | MP-TPT-L | 65.66 | 62.35 |\n\n 3. The ablation study is unconvincing. Why are results provided only on 5 datasets? The proposed methods can lead to performance degradation in many cases, such as in the Flowers102 and Caltech101 datasets. The average performance gain seems to stem from the EuroSAT dataset, which only contains 10 classes and is sensitive.\n\n4. **Effectiveness of design**: The use of random masks on local features as a proxy for random cropping is questionable. I explored this idea in test-time prompt tuning tasks a year ago and found it ineffective, raising concerns about its effectiveness in MP-TPT.\n\n5. **Lack of error bar analysis**: The paper does not include an error bar analysis, which is an important aspect of experimental evaluation.\n\n[1] Task-Oriented Multi-Modal Mutual Learning for Vision-Language Models. ICCV 2023.\n\n[2] Align Your Prompts: Test-Time Prompting with Distribution Alignment for Zero-Shot Generalization. NeurIPS 2023.\n\n[3] Efficient Test-Time Adaptation of Vision-Language Models. CVPR 2024." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. In L107, how can the method enhance inference efficiency when it requires multi-perspective views, which will obviously increase computational and storage costs? Additionally, Table 1 shows that MP-TPT-S has a lower inference time than TPT. What are the different experimental settings between these two methods, and is the comparison fair? Could the authors provide a more detailed analysis of computational complexity and memory usage?\n\n2. The description in Section 3.2.3 is difficult to understand. What is the difference between test time tuning and test time inference? How to generate $\\boldsymbol{f}^{t *}$ and $\\hat{\\boldsymbol{f}}^{t *}$? Additionally, Figure 2c is confusing; how is Eq. 12 applied in Figure 2c, e.g, where is the $\\lambda$ in Figure 2c?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The motivation is clear. Existing test prompt tuning methods focus only on global visual feature augmentation, neglecting the importance of local context in images. By introducing fine-grained local visual features and their corresponding text prompt descriptions, the proposed method should contribute to improved test-time prompt tuning results. The paper is easy to understand." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes utilizing local visual context and class-specific text description augmentation to improve the classification accuracy of the test-time prompt tuning of CLIP model. The local visual representation is obtained by projecting the entire visual feature to the region level and calculating the similarity with text features. The top-K high-similarity region features are selected to produce the class-specific descriptions. The prompts and the global-local visual features are further aligned through a dual interaction during the tuning phase. Experiments show some improvement." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The main weakness of this paper is that the experimental results are marginal. From Table 1, we can see that the best result of the proposed MP-TPT (65.66) is only 0.2% better than the baseline DiffTPT (65.47). Similarly, in Table 2, the MP-TPT method also shows a marginal improvement (less than 0.5%). Did the authors conduct statistical significance tests to verify the effectiveness of the proposed method? These minor differences may also stem from the randomness of the training process. Providing error bars or standard deviations would make the results more convincing. Furthermore, does the method work beyond the CoOp framework, such as on Maple[1] and PromptSRC[2]?\n\n[1] MaPLe: Multi-modal Prompt Learning\n[2] Self-regulating Prompts: Foundational Model Adaptation without Forgetting" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. It would strengthen your claims to include more comprehensive comparisons with a broader range of state-of-the-art methods in your experiments. Highlighting specific scenarios where MP-TPT excels or falls short could provide valuable insights.\n2. Can you clarify the specific roles that global, local, and language perspectives play in test-time prompt tuning? In particular, how do local and language perspectives interact, considering their apparent strong coupling.\n3. Could you provide more experiment on MPTPT+CoOP/MaPLE or other prompt tuning method in Base-to-Novel Generalization? It will help to prove MPTPT’s effectiveness as a plug-to-play prompt learning mthod. \n4. Providing detailed ablation studies that analyze trade-off between speed, accuracy and amount of parameters would enhance the understanding of the practical implications of MP-TPT." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. MPTPT addresses a critical limitation in existing methods that rely solely on global visual features. By incorporating class-specific, region-based prompts, the paper proposes an innovative way to adapt VLMs to unseen data without retraining, which is both effective and practical. \n 2. The methodology is rigorous, with extensive experiments on 15 benchmark datasets that demonstrate the model's adaptability and efficiency, especially in zero-shot and cross-dataset settings. Ablation studies add further credibility by detailing each component's contribution." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces a novel method called Multi-Perspective Test-Time Prompt Tuning (MP-TPT) designed to enhance vision-language models (VLMs) during test time. Unlike prior approaches that focus solely on global visual features, MP-TPT combines global and local visual information with language prompts, offering a comprehensive view during test-time adaptation. The method enhances textual prompts with class-specific descriptions by using local visual information, which allows the model to capture diverse contextual variations. Extensive experiments across multiple benchmarks demonstrate that MP-TPT achieves notable improvements in accuracy and inference speed compared to state-of-the-art methods, particularly in zero-shot and cross-dataset generalization scenarios." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Limited improvement over global feature methods: results indicate that the performance gains of MP-TPT over other methods focusing on global visual features, such as DiffTPT, are not substantial, which raises questions about the effectiveness of incorporating local visuals. \n2. The paper does not sufficiently clarify the interaction between local visual features and text descriptions. A more detailed explanation of how these components integrate during optimization and inference would enhance understanding.\n3. While MP-TPT introduces local visual information to improve class-specific descriptions, the paper could benefit from a deeper analysis of how these local augmentations influence specific categories, particularly when handling complex classes." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024multiperspective,\ntitle={Multi-Perspective Test-Time Prompt Tuning for Global, Local Visuals, and Language},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=0Xc6o1HKXD},\nnote={under review}\n}" }, "abstract": { "value": "Recent advances in vision-language models (VLMs) have demonstrated significant generalization across a broad range of tasks through prompt learning. However, bridging the distribution shift between training and test data remains a significant challenge. Existing researches utilize multiple augmented views of test samples for zero-shot adaptation. While effective, these approaches focus solely on global visual information, neglecting the local contextual details of test images. Moreover, simplistic, single-form textual descriptions limit the understanding of visual concepts, hindering the transfer performance of classes with similar or complex visual features. In this paper, we propose a Multi-Perspective Test-Time Prompt Tuning method, MP-TPT, building on two key insights: local visual perception and class-specific description augmentation. Specifically, we introduce local visual representations from VLMs during the optimization process to enhance the prompts' ability to perceive local context. On the other hand, we design a data augmentation method at the text feature level that imparts regional visual priors to specific class texts, thereby enriching the class-specific descriptions. Furthermore, we synchronize the multi-view concept during the inference, integrating both local and global visual representations with text features for a deeper understanding of visual concepts. Through extensive experiments across 15 benchmark datasets, we demonstrate the advantages of MP-TPT, particularly achieving a 1% improvement in state-of-the-art TPT accuracy in cross-dataset settings, along with 4.5 times acceleration in inference speed." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Prompt Learning", "Test Time Adaption", "Vision-Language Models" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/124bc5767b1aa0b2605b04cdfec0732301de225e.pdf" }, "presentation": null, "primary_area": { "value": "applications to computer vision, audio, language, and other modalities" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/b852656a84e8c244354cdcba614ee69c64efbcfa.zip" }, "title": { "value": "Multi-Perspective Test-Time Prompt Tuning for Global, Local Visuals, and Language" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. Although there is a difference for traditional sign language recognition methods employing such means as MLP and CTC loss, the authors propose for different tasks still use different supervision, for example, words, glosses, and sentences, and why it is still referred to as a unified paradigm.\n\n2. In Fig. 5, why the feature information of the face has to be forwarded to the left Pose Encoder after it has been encoded by the Pose Encoder is not mentioned in the paper..\n\n3. In line 479 of the paper, the authors show a boost of 1.36 on BLEU-4, but the corresponding value is not found in Table 9." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The Uni-Sign framework proposed by the authors utilizes a large-scale generative pre-training strategy and a novel fine-tuning paradigm to bridge the gap between pre-training and downstream sign language understanding tasks in traditional approaches.\n\n2. The Uni-Sign framework achieves significant performance gains on both sign language recognition and translation tasks, and experiments are conducted on multiple datasets.\n\n3. The related work of paper is adequate, investigating research on sign language tasks including pre-training strategies, dataset development, and so on, from a variety of perspectives." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a new pre-training framework that bridges the gap between pre-training and downstream sign language understanding tasks through a large-scale generative pre-training strategy and a novel fine-tuning paradigm that achieves impressive performance in multiple benchmark tests." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The paper is not clear and detailed enough to explain the score-aware sampling strategy, and does not give a detailed analysis of the process or a corresponding explanation in Figure 5, which could lead to potential misunderstandings or errors.\n\n2. The author omitted experimental results on several widely used datasets, such as Phoenix14, Phoenix14T, USTC-SLR 500, USTC-CSL100, etc.\n\n3. As shown in Tables 4 and 6, the proposed Uni-Sign method does not achieve the best performance on multiple datasets of continuous sign language recognition and sign language translation. It even performs worse when more modalities are introduced, which makes me worried about the performance of this work.\n\n4. The number of parameters of the model is not mentioned in the paper. This feedback highlights the importance of including these key performance metrics, as they are critical for evaluating the practicality of the model.\n\n5. It is recommended that the authors make font color changes for the tables throughout the article, due to the large amount of experimental data, while bolding may mislead the reader, especially for Tables 3 through 6." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 2 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please refer to the Weakness section above. If the authors can address these concerns, I would consider raising the rating." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. Uni-Sign effectively unifies multiple SLU tasks, such as isolated sign language recognition (ISLR), continuous sign language recognition (CSLR), and sign language translation (SLT), under a single framework. \n2. The introduction of CSL-News, a substantial CSL dataset, provides a significant resource for the SLU field and addresses the limitations of prior smaller datasets." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes Uni-Sign, a unified pre-training framework for sign language understanding (SLU) tasks, addressing the challenges in existing methods that struggle with transferring knowledge across different tasks. The framework uses a new large-scale dataset, CSL-News, which contains 1,985 hours of Chinese Sign Language (CSL) videos paired with textual annotations. Extensive experiments demonstrate that Uni-Sign achieves state-of-the-art performance across multiple SLU benchmarks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Compared to other datasets, what unique advantages or characteristics does the proposed CSL-News dataset offer besides its longer duration? Additionally, why is the vocabulary size relatively limited, and could the restricted language variety impact pre-training effectiveness?\n2. In the comparisons of downstream tasks in Section 4.3, did other methods also use the CSL-News dataset for pre-training? If not, does this raise any concerns about fairness in the comparisons?\n3. In the comparative experiments, while high-performing results are analyzed, the reasons behind lower performance should also be provided, such as in Tables 4 and 6.\n4. In Tables 3 to 6, what would the results of Uni-Sign be if it used only RGB video?\n5. How do the computational costs, inference time, and memory usage of the proposed model compare to other methods? Does Uni-Sign maintain a competitive advantage in these aspects?\n6. The manuscript includes numerous comparative results, but it lacks visualizations to intuitively demonstrate the model’s effectiveness. More visual presentations for each downstream task are recommended." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "The paper in general addressed the ideas and motivations it introduces. The following question will help add more comprehensive understanding. \nGeneralization and Applicability\n1. Multilingual Evaluation: The sources primarily focus on CSL and ASL. Could the authors comment on the applicability of Uni-Sign to other sign languages? How might the model's architecture and pre-training strategies need to be adapted for multilingual SLU? This is important to assess the generalizability of Uni-Sign and its potential impact on a broader range of sign language communities\n2. Multi-signer Scenarios: How well does Uni-Sign perform in situations involving multiple signers? What challenges might arise in such scenarios, and how could the model be modified to handle them effectively? Addressing this question would provide a more realistic assessment of Uni-Sign's capabilities in real-world applications where multiple signers may be present\n\nComparison and Analysis\n1. Comparison with LLM-based SLT Methods: Recent studies like Sign2GPT and Sign-LLM have explored the use of LLMs for gloss-free SLT. Could the authors provide a comparative analysis of Uni-Sign against these LLM-based approaches? This would help clarify Uni-Sign's contributions and position it within the broader landscape of SLT research\n2. In-depth Analysis of the Unified Fine-tuning Paradigm: How does the shared objective function influence the performance of individual tasks like ISLR and CSLR? Are there any potential task-specific adaptations that could be incorporated within the unified framework to further optimize performance? This analysis would provide a more nuanced understanding of the paradigm's strengths and weaknesses" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "Originality: \n1. The paper presents Uni-Sign, a novel unified pre-training framework for Sign Language Understanding (SLU) that bridges the gap between pre-training and downstream tasks by treating them as a single Sign Language Translation (SLT) task during fine-tuning. This approach deviates from previous methods that relied on indirect pretext tasks or were limited by data scale and transfer capability\n2. The authors introduce CSL-News, a large-scale Chinese Sign Language (CSL) dataset containing 1,985 hours of video with text annotations, considerably larger than existing CSL datasets. his dataset enables effective large-scale pre-training, addressing a gap in CSL resources compared to American Sign Language (ASL) and British Sign Language (BSL)\n3. The paper proposes a Prior-Guided Fusion (PGF) module that utilizes keypoint coordinates as priors to model fine-grained spatial consistency between pose and RGB modalities, going beyond simple spatial-temporal fusion techniques. This approach addresses the representational gap between modalities and leverages keypoints to enhance accuracy. \n4. A score-aware sampling strategy is introduced to address the computational challenges of RGB-pose fusion by selectively choosing RGB frames corresponding to low-confidence keypoints, balancing performance with speed\n\nQuality:\n1. The paper is well-written and presents a clear and comprehensive methodology. The authors provide detailed descriptions of their approach, including data curation, pre-training and fine-tuning strategies, and multi-modal fusion techniques\n2. The ablation studies thoroughly investigate the contribution of each key component, offering insights into the model's performance and the impact of design choices\n3. Quantitative results show that Uni-Sign surpasses previous state-of-the-art methods on multiple benchmarks, including significant improvements in BLEU4 scores for SLT tasks\n\nClarity:\n1. The paper is well-organized and easy to follow.\n2. Figures and tables effectively illustrate the framework, data distribution, and experimental results\n3. Mathematical notations and equations are clearly defined and explained\n4. Qualitative translation examples provide further insights into the model's capabilities\n\nSignificance:\n1. The introduction of the CSL-News dataset addresses a significant need for large-scale CSL resources, potentially fostering advancements in CSL research\n2. The unified pre-training and fine-tuning framework with a generative approach demonstrates a promising direction for improving SLU performance, particularly for SLT tasks\n3. The proposed PGF module and score-aware sampling strategy offer effective solutions for multi-modal fusion and computational efficiency, potentially benefiting future SLU research\n4. The paper's findings have implications for advancing sign language technologies, promoting accessibility and communication for the Deaf/Hard of Hearing community\n5. The authors' commitment to open-sourcing the code and dataset further contributes to the significance of the work, facilitating reproducibility and future research in SLU" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents Uni-Sign, a novel framework for Sign Language Understanding (SLU) that leverages large-scale generative pre-training and a unified fine-tuning paradigm. The paper presents a well-motivated and well-executed approach to SLU. The introduction of the CSL-News dataset and the innovative Uni-Sign framework are significant contributions to the field, demonstrating state-of-the-art performance across various SLU tasks. The paper is well-written and clearly explains the proposed methodology and experimental results. The authors make several notable contributions:\n•Introduction of CSL-News: The authors introduce CSL-News, a large-scale Chinese Sign Language (CSL) dataset comprising 1,985 hours of video-text pairs. This dataset significantly surpasses existing CSL datasets in size and diversity\n•Unified Pre-Training and Fine-Tuning: During fine-tuning, it treats downstream SLU tasks, such as isolated sign language recognition (ISLR), continuous sign language recognition (CSLR), and sign language translation (SLT), as a single SLT task. This unified approach facilitates seamless knowledge transfer and eliminates the need for task-specific fine-tuning methods.\n•Prior-Guided Fusion (PGF) Module: To address the limitations of inaccurate keypoints, the authors propose a PGF module that fuses pose and RGB information using keypoint coordinates as priors. \n•Score-Aware Sampling Strategy: The authors introduce a score-aware sampling strategy to improve computational efficiency. \n•Comprehensive Evaluation: The paper includes a comprehensive evaluation of Uni-Sign across various SLU benchmarks, demonstrating its superior performance in ISLR, CSLR, and SLT tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Discussion on Computational Complexity: While the authors introduce a score-aware sampling strategy to improve efficiency, a more in-depth discussion on the computational complexity of Uni-Sign would be beneficial. This could include analyzing the trade-offs between accuracy and computational cost for different sampling probabilities and exploring potential optimizations.\n2. Further Analysis of CSL-News: While the paper describes the creation of CSL-News, further analysis of the dataset's characteristics, such as vocabulary distribution and linguistic complexity, would be valuable. This would provide a more comprehensive understanding of the dataset's potential and limitations.\n3. Cross-Dataset Generalization: Evaluating Uni-Sign's performance on unseen sign language datasets would demonstrate its generalization capabilities. This could involve fine-tuning the pre-trained model on a different CSL dataset or even a dataset from another sign language, like American Sign Language (ASL). Successful cross-dataset generalization would highlight the robustness of the learned representations and the effectiveness of the unified approach.\n4. Analysis of Error Patterns: A qualitative analysis of the translation errors made by Uni-Sign would provide valuable insights into its limitations and potential areas for improvement. This could involve categorizing errors based on linguistic features, such as sentence complexity, sign ambiguity, or finger-spelling. Identifying common error patterns could guide future research directions.\n5. Exploration of Multi-Signer Scenarios: The authors mention their interest in exploring SLU tasks in complex scenarios, such as multi-signer situations. Including preliminary experiments or discussions on adapting Uni-Sign to handle such scenarios would further enhance the paper's impact and contribution to the field." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 4 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "See above" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "Pros:\n1. This work proposes a unified framework to conduct pretraining and finetuning, which demostrates novelty.\n2. This work shows promising performance across a wide range of benchmarks.\n3. The paper is easy to understand." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper propose Uni-Sign, a unified pre-training framework that eliminates the gap between pre-training and downstream SLU tasks through a large-scale generative pre-training strategy and a novel fine-tuning paradigm. It also introduce CSL-News, a large-scale Chinese Sign Language (CSL) dataset containing1,985 hours of videos paired with textual annotations." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Questions and cons:\n1. During the data curation process, the authors use a ASR toolkit (FunASR) to convert the speech into texts as labels. There exist some problems. First, as the speech signal has time delay with the sign language expressed by the signer, how to assure that the temporally cropped clips are exactly aligned with the transcribed texts? Second, the authors have stated the averaged length of words are 40 and the averaged length of clips are 9.5s. It's very hard to express 40 words within 9.5s for a signer. Thus, it's most probably that the signer has neglected some meanings in the sentence, and only expressed a part of mearnings in the signs. In this condition, the signs are probably not aligned with the transcribed texts. Third, i observed that in the paper, the authors don't organize a double-check process for the cropped videos from the TV shows to check the alignment between texts and clips, the correctness of transcribed texts, the correctness of transcribed signs and other things. Thus, how to assure the compleness and correctness of the curated datasets?\n2. During the experiments for CSLR, PHOEXNI14 and PHOENIX14-T are also broadly used datasets. Why not report the results on these datasets? It's due to the language gap between pretraining data and downstream data? How about the performance on these two datasets?\n3. In table 3 and table 5, some other numbers are bolded except the results reported by the proposed method. The authors may clarify on this or use another way to emphsize the results." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "In Table 6, the performance of the proposed method is somewhat lower than the existing baseline SSVP-SLT. Although it is not a very big issue to me, but I would like to know more about it. Why is this the only (rather large) SLT dataset where the proposed method achieves sub-optimal results?\n\nThe ablation results shown in Table 7 are rather strange as compared to Tables 8-10, because the settings are different. Table 7 runs experiments on ISLR and CSLR, Tables 8-10 run experiments on CSL-Daily for CSLR and SLT. Why are these different? Moreover, Table 7 and 8 are run in the pose-only setting while Tables 9 and 10 are in the RGB-Pose setting, why should this be the case?\n Furthermore, some of the more important experiments (Table 7 and 8 in my opinion) should be evaluated on all three different sign language understanding tasks.\n\n\nWhat is the impact of the pre-training? This crucial aspect as not been evaluated properly. For instance, what if the model is trained only using the fine-tuning stage (Stage 3), but for a longer time (i.e., matching the overall training time of the pre-train then fine-tune approach)? How does this affect the performance? This is important as it shows us the benefits of pre-training. Although some results have been provided in table 7, the results and implications are not clear to me. Furthermore, the task-specific training settings and details have not been mentioned." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "Developing a unified approach to handle the various sign language understanding tasks is meaningful. In some sense, the work extends some recent LLM-based sign language understanding works by including the aspect of unifying across the sign language understanding tasks. \n\nThe authors introduce a new large-scale sign language dataset for Chinese Sign Language. This dataset could be quite useful for further progress in the field.\n\n\nThe experiment results are quite impressive, especially on the gloss-free SLT task. In my opinion, gloss-free SLT is the setting that is the closest to real applications, so this is quite good." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper has two main contributions. 1) A Uni-Sign method for tackling the three sign language understanding tasks in a unified manner. The model first pre-trains on a large sign language dataset via language modeling, then is fine-tuned on each of the individual tasks separately. 2) A CSL-News dataset, which is a large-scale Chinese Sign Language dataset. Some other minor architectural designs are also proposed. Overall, the proposed method performs quite well across the three sign language understanding tasks, and particularly performs well in Sign Language Translation." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The proposed method is not very novel. The proposed pre-training approach is to train the model in a language modelling manner, while also using visual features from the sign videos. Then, for the fine-tuning, the language modelling loss is again used for the various tasks. There are some minor contributions, such as a prior-guided fusion module and a score-aware sampling strategy, but these do not seem quite so substantial.\n\nI think that in the related works discussion, there should be a part discussing some other works in other fields employing language modelling (or sequence modeling) for tackling various tasks in a unified manner. For instance, this has been done for image-based tasks, and may have also been done for pose-based tasks. This will give the reader a better understanding of the developments of the “unifying via language modeling” paradigm.\n\n\nMore specific concerns and questions are in the “Questions” section." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024unisign,\ntitle={Uni-Sign: Toward Unified Sign Language Understanding at Scale},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=0Xt7uT04cQ},\nnote={under review}\n}" }, "abstract": { "value": "Sign language pre-training has gained increasing attention for its ability to enhance performance across various sign language understanding (SLU) tasks. However, existing methods often suffer from a gap between pre-training and fine-tuning, leading to suboptimal results. To address this, we propose Uni-Sign, a unified pre-training framework that eliminates the gap between pre-training and downstream SLU tasks through a large-scale generative pre-training strategy and a novel fine-tuning paradigm. First, we introduce CSL-News, a large-scale Chinese Sign Language (CSL) dataset containing 1,985 hours of video paired with textual annotations, which enables effective large-scale pre-training. Second, Uni-Sign unifies SLU tasks by treating downstream tasks as a single sign language translation (SLT) task during fine-tuning, ensuring seamless knowledge transfer between pre-training and fine-tuning. Furthermore, we incorporate a prior-guided fusion (PGF) module and a score-aware sampling strategy to efficiently fuse pose and RGB information, addressing keypoint inaccuracies and improving computational efficiency. Extensive experiments across multiple SLU benchmarks demonstrate that Uni-Sign achieves state-of-the-art performance across multiple downstream SLU tasks. We will release the source code and the dataset to the public." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Sign language understanding", "Pre-training", "Large-scale sign language dataset" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/c74b74b55de2008288081544ead353474a559f53.pdf" }, "presentation": null, "primary_area": { "value": "unsupervised, self-supervised, semi-supervised, and supervised representation learning" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Uni-Sign: Toward Unified Sign Language Understanding at Scale" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "- The reasoning behind the name 'MMComposition' is unclear." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- This paper presents a comprehensive benchmark focused on compositionality, encompassing a wide range of skills from perception and reasoning to probing.\n\n- This paper provides an extensive evaluation of recent models, including both open-source and API-based models, highlighting areas where they continue to fall short of human capabilities.\n\n- The paper is well-written with clearly organized sections." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces MMComposition, a QA benchmark that evaluates the compositional capabilities of modern vision-language models. MMComposition encompasses a range of tasks, including perception, reasoning, and probing, with multiple subtasks presented in various QA formats: yes/no, multiple-choice, and indefinite-choice. The dataset is curated from numerous existing sources, with QA pairs annotated by humans. Covering 13 distinct vision-language compositionality tasks, this benchmark offers a comprehensive evaluation of both proprietary and open-source vision-language models. The paper also analyzes factors that may influence the compositional abilities of VLMs, such as the resolution of visual encoders, the scale of language decoders, and the volume of training data." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Although the benchmark includes diverse skill sets and QA formats, the specific aspects that pose challenges are not clearly defined. It is also unclear what distinguishes this benchmark from other general QA datasets designed to test modern VLMs for AGI, such as MMMU, MMStar, and tons of similar benchmarks. The paper does not provide comparisons in terms of general capabilities across QA datasets; instead, it focuses on embedding-based benchmarks for comparison, as shown in Table 1. Comparing the scale of evaluation samples, such as the number of images or questions across different benchmarks, would also be valuable.\n\n\n- Related to the first weakness, one might question whether this benchmark is truly challenging. Some compositionality benchmarks or visual QA tasks could potentially be solved using only language models in an image-blind setting, due to language priors, such as coherence, grammar, and clues embedded across answer choices. As specific example, in second example in figure 3, can is often made of metal, such knowledge aids in answering correctly without relying on visual cues. It would be beneficial to examine the proportion of questions that can be solved solely using large language models. \n\n\n- Several essential details are missing regarding the benchmark construction. In the human annotation process, additional information is needed: Who annotated the dataset? How was confidence measured, and how were errors handled in finalizing the annotations? Additionally, it’s unclear how misaligned captions were manually added in the probing task (line 255). Furthermore, for reporting human performance, what was the process? It would be important to present individual human performance scores for each skill, rather than a single overall score.\n\n\n- The empirical trends concerning the scale of the visual encoder, language decoder, and training data are perhaps not surprising. The paper does not analyze whether these trends are specific to the proposed benchmark or if they also appear in other general visual QA benchmarks. Meanwhile, an additional suggested analysis could explore how the design of the visual connector (e.g., fully connected layer or Q-Former style) and the method of visual token insertion (e.g., tokens input directly into the language model or through cross-attention connections) impact performance of the proposed benchmark. \n\n\n- There are some notable clarity issues, including typographical errors such as 'MuriBench' in line 237 and 'ARC' in line 241. Additionally, there are inconsistencies in publication years for certain cited papers, particularly recent NeurIPS papers like SugarCrepe, which collectively raise concerns about professionalism. \n\n\n- Could fine-tuning VLMs on specific datasets improve performance on MMComposition?\n\n---\n\nAssessment: While the extensive evaluations across VLMs are commendable, the benchmark falls short of expected standards in terms of detailed documentation, verification, and comparisons with other QA benchmarks. Additionally, analyses of the proposed benchmark could be enhanced by comparing observed trends with those from other benchmarks." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "humans involved in data curation, without reporting details on that, however I am not sure if this is a real ethical concern here" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "please see weaknesses" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "* new benchmarks are always good, human curation is appreciated\n* large number of models evaluated" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a new compositional reasoning benchmark that is constructed from existing benchmarks (data collection, lines 211-238) augmented with negative options retrieval by similarity, consensus filtering by several recent LMMs and further human filtering of the resulting visual QA. An extensive evaluation of recent models on the proposed benchmark is performed. Some additional ablations are attempted by grouping models trained on more data, larger LLM decoders, vis. encoder combinations etc. However, those only confirm known facts: larger data, larger decoders, or more encoders are beneficial. Some analysis of failures is provided, albeit only qualitative. Main interesting aspect seems to be a large gap reported between human performance and the models. However, no statistics of the human subjects are provided (eg how many humans were employed, how they were motivated, what was the disagreement between humans, age groups, etc.)." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* no new data is introduced, built from existing benchmarks\n* no surprising conclusions\n* no statistics for the human subjects\n* error analysis is only qualitative\n* dataset construction methodology involving humans could be more interesting - eg. humans could generate questions, red-team models to generate hard negatives etc.\n* I disagree with Table 1, many benchmarks, including those listed have fine-grained questions, there are benchmarks (eg NVLRv2) involving multiple images, other benchmarks have human filtering, at least a partial subset, the only thing I indeed did not encounter before is \"multiple right answers\" (indefinite choice) - which could indeed be a contribution of the paper\n* while benchmark contributions are appreciated, it seems this paper is somewhat below what I would expect from the level of contribution of an ICLR paper" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "see weaknesses" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "Strengths of MMCOMPOSITION:\n\n1. Targeted Evaluation of Compositionality for VLMs: MMCOMPOSITION provides a focused benchmark to assess compositional reasoning in Vision-Language Models, an area where existing models often fall short. By going beyond basic attribute recognition, MMCOMPOSITION evaluates tasks like multi-image reasoning, object interactions, and counting, all of which are crucial for real-world, nuanced understanding.\n\n2. Improvement upon Existing Compositional Datasets: This benchmark builds on and enhances data from existing compositional datasets, such as ARO, to create a more diverse and challenging evaluation framework. By curating tasks that move beyond traditional benchmarks, MMCOMPOSITION offers a comprehensive dataset for testing complex visual-language interactions.\n\n3. In-Depth Model Comparison and Component Analysis: MMCOMPOSITION evaluates over 50 VLMs across different architectural components, allowing a detailed comparison. This thorough assessment reveals how factors like encoder resolution, decoder size, and training data diversity impact compositional reasoning. It offers practical insights that can guide future improvements in model design." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper \"MMCOMPOSITION: Revisiting the Compositionality of Pre-Trained Vision-Language Models\" presents MMCOMPOSITION, a new benchmark focused on testing VLMs' ability to handle complex compositional tasks like object interactions, counting, and scene reasoning. With 4,342 annotated questions across 13 categories, the benchmark highlights a clear performance gap between models and humans (67.95% vs. 90.31% accuracy). Results suggest that improving high-resolution encoders, scaling language decoders, and expanding training data are key to better compositional reasoning in VLMs. MMCOMPOSITION offers a practical tool for refining future VLMs to better understand complex compositions." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "typos:\ntable 4 - Relolution \n\n1. In-context multimodal compositionality: Adding tests for in-context multimodal compositionality could strengthen the benchmark, as this capability is crucial for real-world applications. Evaluating models' ability to maintain compositional understanding across multi-modal inputs, rather than isolated tasks, could enhance the dataset's relevance.\n2. Multi-hop compositional problems: The paper would benefit from including multi-hop reasoning tasks, where models must integrate multiple compositional steps to arrive at an answer. This kind of problem is essential for advanced compositionality and would make the benchmark more challenging and comprehensive.\n3. Questionable novelty: The novelty of the paper could be improved if it incorporated points 1 and 2. Adding in-context multimodal compositionality and multi-hop compositional problems would make MMCOMPOSITION a more distinctive and valuable benchmark." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1) In line 254, “we select several captions from the dense captions in Visual Genome as the correct options and write the misaligned captions manually for the image”\nWhat are the criteria for writing the misaligned captions? In terms of which characteristics do the misaligned captions differ from the original captions?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1) The dataset is human-annotated and covers a wide range of tasks in terms of compositional understanding\n2) The paper evaluates 54 representative large VLMs including open-source and proprietary ones. The benchmark is challenging and demonstrates large performance gap between human and VLMs. \n3) The analysis on model component provides valuable insight on model design." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes MMComposition - a human-annotated benchmark dataset for evaluation of the compositionality of large Vision-language models. \nThe benchmark contains 4.3K questions in three main dimensions: perception, reasoning and probing which are divided into 13 categories. There are both questions that contain a single image and multiple images. Most questions have a single correct answer. There are 459 questions with indefinite-choice.\nThe benchmark demonstrates human performance (90.31%) and state-of-the VLMs (best performance of 67.95% among 54 evaluated VLMs). \nThere is also analysis of impact of VLM architecture factor on the benchmark performance, e.g. visual encoder design, language decoder size, training data volume." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1) The paper categories the questions into 4 difficulty levels based on the performance of 6 open-source models. In Figure 8, it shows that 62.09% of questions in the category “superhard” lead to the average performance on all VLMs below the average level. It would be interesting to analyze what characteristics lead to the different difficulty levels of these questions? This can shed light on how to design difficult questions for the competent VLMs. \n2) In the evaluation benchmark, for questions that contain multiple images, the images are concatenated into a big collage and fed into the model. Some of the VLMs have multiple-image samples in the training data and can perform VQA with multiple input images. Does it impede the performance of these models to feed the collage into them?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024mmcomposition,\ntitle={{MMCOMPOSITION}: Revisiting the Compositionality of Pre-trained Vision-Language Models},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=0YXckVo7Kw},\nnote={under review}\n}" }, "abstract": { "value": "The advent of large Vision-Language Models (VLMs) has significantly advanced multimodal understanding, enabling more sophisticated and accurate integration of visual and textual information across various tasks, including image and video captioning, visual question answering, and cross-modal retrieval. Despite VLMs' superior capabilities, researchers lack a comprehensive understanding of their compositionality -- the ability to understand and produce novel combinations of known visual and textual components. Prior benchmarks provide only a relatively rough compositionality evaluation from the perspectives of objects, relations, and attributes while neglecting deeper reasoning about object interactions, counting, and complex compositions. However, compositionality is a critical ability that facilitates coherent reasoning and understanding across modalities for VLMs. To address this limitation, we propose MMCOMPOSITION, a novel human-annotated benchmark for comprehensively and accurately evaluating VLMs' compositionality. Our proposed benchmark serves as a complement to these earlier works. With MMCOMPOSITION, we can quantify and explore the compositionality of the mainstream VLMs. Surprisingly, we find GPT-4o's compositionality inferior to the best open-source model, and we analyze the underlying reasons. Our experimental analysis reveals the limitations of VLMs in fine-grained compositional perception and reasoning, and points to areas for improvement in VLM design and training." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Vision-Language Models", "Compositionality", "Benchmark" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/1c1bc8887e836fe08a57ac7d15ee1efa86b90656.pdf" }, "presentation": null, "primary_area": { "value": "datasets and benchmarks" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "MMCOMPOSITION: Revisiting the Compositionality of Pre-trained Vision-Language Models" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. what is \"t \\wedge T\" ?\n 2. sec 3.3, what is P(s_t) a posterior over?\n 3. In what sense is \\pi_t a \"potential\" function?\n 4. I cannot make any sense of eq 2. Is w \\in V the same as w_t? why cant you simply remove the indicator function and write it as \\sum_w \\in C ? why is the indicator function in the deminator as well? is the intent to have a logit distribution that only puts mass on the tokens in C?\n 5. are the rollouts done on the backbone model or the expert model? have we considered /measured the inference time cost? this is an important consideration in a paper about mtcs type methods.\n\n6. Does q_\\phi simply reward completions of the output that have tokens in C?\n\n7. intro: \"in contrast, an expert model.....\" : this is an interesting claim (does seem plausible). is there a citation for evidence?\n\n8. line 140: tend to generate similar token....\": what does this mean?\n\n9. i am not up to date on the faithfulness literature, but the kind of interventations that the paper describe as standard ways of evaluation i.e. word inclusion and perturbation just seem to be likely to be noise-prone, leading to unreliable evals?\n\n10. GenExpert =? lookahead?\n\n11. comment: the discussion between 329-342 helped understanding a bit and should be earlier in the paper.\n\n12. sec 5.2.2: the NLI example is a bad one i think. Submergible only means it is something that can be submerged. which doesnt automatically mean it is submerged." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "1. MTCS type inference is a hot topic right now, and it is indeed an important frontier for LLMs to improve on.\n2. At a surface level, experimental results seem to show large gains.\n3. There" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes an approach to do faithful rationale generation in LLMs. It uses a steering-based approach to make the outputs more faithful to the reasoning of the llm in classification. The idea is to weight token logits using 2 kinds of reward models: A \"local\" one that tries to match tokens to those suggested by a domain-specific expert model and a \"lookahead\" one that does an MTCS type search and re-weights logits based on rewards from unrolled sequences. \n\nExperiments are performed on a couple of QA type datasets, demonstrating that each method makes improvements in classification accuracy and faithfulness of rationales. Some qualitative analyses are also presented." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Section 3 is pretty badly written, it is pretty hard to get the details of the approach. Instead of invoking irrelevant sophisticated-sounding terminology like \"Feynman-Kac\" formulas it would be better to describe the method in more detail. The math especially is confusing, see below. \n\nThe paper seems to show some positive experimental results, but I am concerned about whether we are looking at a meaningful comparison. The proposed methods rely on domain experts. Looking at table 8 in the appendix these are generally models that have been fine-tuned for the task in some way (and not just on the validation sets as the main section claims, some have access to external datasets). So it shouldn't be that surprising that a method that is given access to an expert which has more signal will do better than the backbone pre-trained model. A fair comparison would have to be with an approach that does vanilla fine-tuning of LLama or mixtral model. \n\nIn terms of novelty: The authors have not really cited relevant work in the controlled decoding space:\n\nhttps://arxiv.org/abs/2310.17022\n\nhttps://sea-snell.github.io/ILQL_site/\n\nThese works already do something more sophisticated than just token reweighting by a reward score. So what is the novel contribution here? 2 possibities:\n1. Focusing on the faithfulness problem.\n2. The \"lookahead\" idea of the reward model. I dont recall having seen this before, but it feels like a simplification of a full-blown MCTS. I would also call this a poor man's version of ILQL.\n\nSo we are just left with #1 then, unless I missed something. And this is something I consider of limited novelty (more like an application for a particular problem, though one with interesting implications from the steering perspective)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "Could this method generalize to the setting where the rationale is generated before the answer?" }, "rating": { "value": 1 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. Faithful rationales are important for explainability and model control, which makes this work well-motivated.\n2. The proposed method is training-free (although with reliance on trained expert models), making their method portable.\n3. A comprehensive set of experiments is conducted to showcase the effectiveness of their proposed method." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The work aims to improve the faithfulness of the LLM-generated rationales for reasoning tasks. They propose an inference-based method where an LLM is guided to generate more faithful rationales by both local and global rewards. Both rewards are provided by additional expert models which are trained on the downstream tasks. Experiments demonstrate the effectiveness of the method in achieving higher accuracy and faithfulness." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The method requires the model to generate the answer prior to the rationale, which provides no guarantee that the decision is made based on the rationale. The model could still suffer from inherent biases.\n2. The method is limited to reasoning tasks with constrained answer space, limiting its generalization to more open-ended tasks.\n3. The method is poorly introduced. It would be very helpful if the authors could explain what exactly Eq.1-3 are doing in plain words." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "1. If the expert model is as big as the base model, how can the computational cost be similar to beam search?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The direction this paper explored has been receiving increasing interest recently: improving the quality of LLM answers at inference time without modifying the model weights directly. The proposed method improves the zero-shot accuracy and faithfulness of two strong general instruction-tuned models (Llama-3-8B and Mistral-7b-Instruct-v0.3) on three reasoning tasks. The experiment showing the benefits of going beyond local/token-level rewards and taking into account the global/lookahead reward is interesting." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work proposes an inference-time method to improve the performance and faithfulness of general (instruction-tuned) large language models (LLMs). Specifically, the method uses expert models to provide fine-grained and lookahead rewards to search and reweight possible tokens or continuations proposed by the LLM. With the help of expert models trained on the target task or domain, the proposed method can improve both the accuracy and faithfulness of the zero-shot answers of two instruction-tuned models on three reasoning tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- There needs to be more details explaining the proposed method, the motivation of each part, the equations and variables, the relation to related work, and the implementation details. Specifically:\n - Section 3.3: how does the Feynman-Kac Formulae model inspire the faithfulness-seaking search framework? The connection is not straightforward. The notation of eq 1 is ambiguous. What does posterior P_t(st) mean exactly? How is it used in the proposed method? Also, the equation itself needs more explanations on what it is computing and why in this way.\n - Section 3.4 (Local constraint): line 179 I find it hard to follow the motivation. How \"certain attributes can be implicitly conveyed over longer spans rather than the individual token\" is connected to \"Instead, domain-specific experts tend to demonstrate better accuracy in knowledge-rich tasks.\"? If the domain expert has better accuracy why not just use the expert to predict the scores? Why bother to use them to improve the backbone LLM? In lines 180-181, it says \"we introduce a set of classification label words C from these expert models ...\", how is C constructed? What is the motivation behind token masking?\n - Section 3.4 (Lookahead Reweight): Equation 3 is hard to understand without proper explanations. $m$ and $x_i$ are not explained in the texts. $s_{t+l}=s_{t-1}||w_t$ is more confusing: $s_{t+l}$ has $t+l$ tokens while $s_{t-1}||w_t$ has $t$ tokens. What does equality mean here?\n- Many experimental details are missing, and important experiments are missing.\n - Missing baselines: the performance and faithfulness of the expert models alone. If the faithfulness or accuracy of the expert models are better than the backbone LLM, why do we even need to use the expert models to improve the backbone LLM?\n - Evaluation details: how is the original model evaluated? If it is a zero-shot evaluation. What is the exact prompt and task format used? How to extract answers from the outputs to calculate the accuracy? The backbone LLMs are state-of-the-art instruction-tuned models. However, the task performance as well as the faithfulness are quite low, so the authors need to provide more details on the evaluation.\n - What is the choice of hyperparameter n (number of rollouts) and how is it chosen?\n- The writing of the paper could be improved for better readability. First, the paper is not properly scoped. For example, in lines 16-18, it says \"... to ensure that LLM-generated rationales are logically coherent and comprehensive.\" However, there is no result discussing the logical coherence or comprehensiveness of answers in the paper. Another example is line 108: it says \"We firstly introduce the faithfulness definition in our context,\", but there is no clear definition in section 3.2." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. The overall experiments are conducted on LLaMA3. I think more backbone LLMs and other sizes of LLMs are needed to justify the proposed inference paradigm.\n2. More experiments on more related datasets is needed." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The paper introduces a novel probabilistic inference method with a dual-reward mechanism, combining local and global reward. This is a very novel solution. \n2. The paper is well-written. I am not an expert in this domain but I can get their core contributions. \n3. The experiment design is clear: they design the ablation study in Section 5.1 to justify the local and global rewards for the final performance. Although I suggest authors could do better by choosing more LLMs in different model size to better support their experimental design." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "They tackle the rationale generation tasks in LLMs' reasoning process. Specifically, they propose a probabilistic inference paradigm that provides fine-grained and lookahead rewards to instruct LLMs to generate good rationale. The key problem addressed is that LLMs often produce unfaithful explanations, especially when they fail to incorporate essential contextual information. \n\n+ **Local Reward**: this component ensures coherence with the immediate context, often by using a domain-specific expert model.\n+ **Global reward**: This assesses the plausibility of the current token in relation to desirable future attributes\n\nThe search algorithm, especially for lookahead reweight seems interesting.\n\nPlease forgive me if I misunderstand something. I spent much time for reading the paper but to be honest, I am not an expert in this area. I will available on the rebuttal time for author's response and will read their response. I am also open to other reviewers' opinions." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. There are several related works that are missing or less discussed: \n + Evaluating Human Alignment and Model Faithfulness of LLM Rationale\n + On Measuring Faithfulness or Self-consistency of Natural Language Explanations\n2. Figure 2 about the distribution of domain-specific words is unclear to me. \"showing that our method can respond more actively to those domain-specific words\" Why does this part matters to the experimental results." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose a probabilistic inference paradigm that provides fine-grained and lookahead rewards to ensure that LLM-generated rationales are accurate and faithful.." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024enhancing,\ntitle={Enhancing {LLM} Faithfulness in Rationale Generation via Dual-Reward Probabilistic Inference},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=0Yfjerm9Zp},\nnote={under review}\n}" }, "abstract": { "value": "As large language models (LLMs) are increasingly applied to complex reasoning tasks, achieving both accurate task performance and faithful explanations becomes crucial. However, LLMs often generate unfaithful explanations, partly because they do not consistently adhere closely to the provided context. Existing approaches address this problem either rely on superficial calibration, such as decomposed Chain-of-Thought prompting, or require costly retraining to improve model faithfulness. In this work, we propose a probabilistic inference paradigm that provides fine-grained and lookahead rewards to ensure that LLM-generated rationales are logically coherent and comprehensive. These rewards are derived from a domain-specific proposal distribution, allowing for optimised sequential Monte Carlo approximations. Our evaluations across three different reasoning tasks show that this method, which allows for controllable generation during inference, improves both accuracy and faithfulness of LLMs while keeping computational costs similar to those of existing decoding techniques. This method offers a promising path towards making LLMs more reliable for reasoning tasks without sacrificing performance or efficiency." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "interpretability", "faithfulness", "Large language model", "constrained generation" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/12e8d9072c1ffa040599a111ca19715109539a78.pdf" }, "presentation": null, "primary_area": { "value": "interpretability and explainable AI" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Enhancing LLM Faithfulness in Rationale Generation via Dual-Reward Probabilistic Inference" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "In conclusion, the paper presents an interesting idea, but the experimental section needs significant refinement. Adding more comprehensive experiments and ablation studies would strengthen the conclusions and clarify the potential of this approach." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The approach is promising; using a generator to produce new samples is a valuable innovation for improving active learning systems. This strategy assumes a pre-trained generative model, which is reasonable for text but may not be universal across domains. The selection criterion is sensible, and directly training the generator to maximize it through RL is more robust than simple thresholding." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper addresses active learning by leveraging a generative model to produce unlabeled examples, which are then labeled by an oracle and added to a classification model's training set. Unlike traditional methods that rely on a fixed pool of unlabeled data, this approach actively generates new, potentially more informative examples. The model prioritizes examples based on their distance from nearest neighbors and the discrepancy in predictions between the generated sample and its neighbors. To guide the generative model in producing high-quality samples, it is trained via a Reinforcement Learning algorithm (PPO), optimizing it to generate samples that best serve the classification task. The method is tested on text classification problems, showing mixed results compared to current state-of-the-art techniques." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "However, the experimental section lacks detail to fully evaluate the approach. Key hyperparameters—such as the number of samples generated per iteration and PPO settings—are not systematically analyzed, and no ablation study is provided. It would also be valuable to see a comparison of results with and without the RL approach. The current experimental section leaves significant space unexplored, making it hard to discern the model’s strengths and weaknesses.\n\nIt’s also unclear how the RL component is applied: Is the policy trained concurrently with sample generation, or is it established before the active learning phase? If the reward function evolves as new samples are generated, this could introduce non-stationarity, which would impact performance. Further clarification on this point is essential.\n\nRegarding performance, the results do not clearly outperform existing methods. Notably, the learning curves for the proposed method (Signal) appear to extend longer than others. This might be because other methods are restricted to samples in the original dataset, while Signal can generate an infinite number of examples. However, this is not entirely clear, as baseline methods don’t achieve fully supervised performance, which raises questions about their comparison criteria." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "As discussed in the weakness section." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The method addresses the limitations of traditional pool-based methods by generating informative synthetic data instances, this could be beneficial when even unlabeled data is scarce.\n2. The paper is mostly well-organized.\n3. The acquisition function that combines both informativeness and relevance makes sense." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces Self-Informed Generative Active Learning (SIGnAL), a RL-based approach for query-synthesizing active learning. SIGnAL generates synthetic data instances to enrich the data pool, especially when access to diverse, unlabeled real data is limited. Experimental results show SIGnAL’s performance advantage over traditional pool-based methods in text classification tasks, particularly when the data pool is very small." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The proposed SIGnAL does not generate the most informative/beneficial data point for labeling, instead, it still requires traditional acquisition function to make the selection. I think this is a critical weakness to this paper. From my understanding, generative AL should not only generate data samples, but more importantly generate the most informative samples.\n2. The settings of this paper is king of niche, most areas that benefit from AL have abundant amount of unlabeled data, if SIGnAL simply generates more unlabeled data, I don't see it being very useful in practice.\n3. The acquisition (relevance and informativeness) is quite simple, relevance is simply the distance, with informativeness directly taken from CAL.\n4. The experiments are very limited. The only results are in Figure 3, with limited datasets, baselines, and the improvements are hardly distinguishable in my opinion. \n\nIn general I think this paper presents an interesting direction, but the details needs a bit more refinements." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "In the equation of line 186, the distribution shouldn't be p_z because there is synthetic data while p_z is defined as real data, right?\nLine 284: missing space between the and generate" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "The idea is interesting and novel, the paper is easy to follow." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes the Self-Informed Generative Active Learning (SIGnAL) framework, which generates synthetic data to improve active learning when real data is scarce. Using reinforcement learning, SIGnAL’s generator produces informative data guided by a reward system, ensuring relevance and usefulness for model training. An acquisition function assesses this data’s informativeness and relevance, optimizing the generator’s outputs. Experiments on text classification validate SIGnAL’s effectiveness, particularly in data-limited scenarios, offering a cost-efficient solution." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The experiments are far from sufficient for a top-tier conference, now there is only overall performance but lack of ablation study and analysis.\n2. As a method that combines active learning and synthetic data generation from LLM, the authors only compare it with active learning approaches, I think they should also compare the proposed method with synthetic data generation without active learning\n\nMissing related work:\n\n[1] Large Language Model as Attributed Training Data Generator: A Tale of Diversity and Bias\n[2] ZeroGen: Efficient Zero-shot Learning via Dataset Generation\n[3] Generating Training Data with Language Models: Towards Zero-Shot Language Understanding" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "The PPO reinforcement learning method is utilized in this paper to optimize AL strategies for larger rewards. Could you provide a detailed explanation of the state and action settings in this RL scenario? \n\nAdditionally, when considering other RL-based active learning strategies, is it worth considering adopting the classifier's accuracy as an additional reward after generating samples?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "By utilizing reinforcement learning, the approach effectively addresses the challenges posed by the dynamic and delayed nature of informativeness, treating instance informativeness as the reward signal to optimize the generative model. The method incorporates an acquisition function that evaluates both traditional informativeness and the relevance of data instances, transforming these evaluations into rewards during training." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper leverages a RL policy to guide data generation, allowing the generator to receive rewards that encourage the creation of more informative instances. Additionally, it introduces an acquisition function that evaluates both informativeness and relevance, seamlessly transforming this evaluation into rewards for optimizing the generator." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The paper provides a detailed analysis of the challenges faced by pool-based active learning methods; however, it lacks an introduction to existing query-synthesizing methods and a distinction between the proposed method and existing synthesizing-based methods, such as “LADA: Look-Ahead Data Acquisition via Augmentation for Deep Active Learning” and “When Active Learning Meets Implicit Semantic Data Augmentation”. However, synthesizing-based methods are one of the primary categories in the active learning scenarios.\n\nThe PPO reinforcement learning method is utilized in this paper to optimize active learning strategies for larger rewards. Could you provide a detailed explanation of the state and action settings in this reinforcement learning scenario? Additionally, is it worth considering adopting the classifier's accuracy as an additional reward after generating samples?\n\nRegarding the experiments: The baseline methods adopted in the paper are all pool-based active learning methods. To further validate the effectiveness of your method, it is suggested to compare with synthesizing-based methods as well. Moreover, according to the experimental setup, synthesizing-based methods annotated twice as much more data, which could account for their superior performance. It is recommended to include ablation studies to provide additional explanations." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. What is the ratio of generated samples to actual unlabeled data queried at each iteration in the active learning process? If generated samples are only queried after unlabeled data, their value seems minimal.\n2. How were the generated samples human-labeled? It’s likely some generated samples are incoherent, making labeling challenging. The paper includes no information about the human labeling process." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The proposed approach outperforms other techniques, such as CAL, BERTKIM, and BADGE.\n2. It allows performance gains by querying generated data after exhausting unlabeled data." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces an active learning approach for NLP tasks utilizing a generative model. It incorporates KL divergence, as proposed in CAL, to retrieve informative samples and uses inter-sample distance to avoid querying unrelated samples. The method outperforms comparable approaches and can continue the active learning process even without access to further unlabeled data by leveraging generated samples. However, the limited datasets and reliance on LLM raise questions about the necessity of the approach." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The utility of this approach is ambiguous. Active learning aims to efficiently query valuable samples in low-data regimes, particularly in areas with difficult labeling requirements, such as medical or legal fields.:\n- 1.1. The paper only presents general datasets (SST-2, AGNEWS, QNLI) focused on tasks like sentiment analysis and topic classification, where active learning might be unnecessary. Given that the LLM itself can achieve higher performance on such tasks, training an additional classifier via active learning seems contradictory. For this approach to be useful, the active learning-trained model should outperform the LLM.\n- 1.2. In this regard, domain-specific datasets, such as PubMed or legal datasets should be added. However, studies suggest that even specific tasks can achieve performance gains without active learning (or human labeling) through LLMs [1], raising questions about this method's utility compared to such approaches.\n2. The number of datasets and class diversity are limited, with only three datasets and two or four classes per dataset. Include datasets with more classes, like DBPEDIA with 14 classes, to address whether the proposed method benefits persist as class counts increase.\n3. The main paper lacks a definition of $\\Phi$, which can only be inferred as a text encoder model.\n4. No ethics statement is provided. An ethics statement and societal impact are mandatory for ICLR.\n5. Hyperparameters are not disclosed. Without code submission, at least hyperparameter settings or a code statement should be included.\n6. Time consumption details are missing. Given the method's reliance on LLMs and RL and the continuous dataset expansion, it likely requires considerably more time than alternative methods. Please add this information.\n\n[1] Kim et al., \"SELF-EXPERTISE: Knowledge-based Instruction Dataset Augmentation for a Legal Expert Language Model\"" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose the Self-Informed Generative Active Learning (SIGnAL) framework which actively generates and selects data instances for annotation and downstream model training." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024selfinformed,\ntitle={Self-Informed Generative Active Learning},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=0YkZe9nwiC},\nnote={under review}\n}" }, "abstract": { "value": "Active learning has been a cost-efficient approach to obtaining high-performance AI models with fewer selective annotations. In scenarios where the acquisition of original unlabeled data poses significant challenges, active learning harnessing synthesized data instances is more promising than traditional pool-based methods. In this paper, we propose the Self-Informed Generative Active Learning (SIGnAL) framework as an effective solution to actively generate and select data instances for annotation and downstream model training. In SIGnAL, we propose to guide the data generation based on a reinforcement learning policy, where the generator is self-informed by the reward to generate more informative instances. In addition, we introduce an acquisition function that measures both the informativeness and relevance of instances. Such acquisition function can be transformed to the reward seamlessly for generator optimization. Our experiments on the text classification task validate the effectiveness of our framework, especially when the original data scale is limited." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Active Learning", "Large Language Model", "Synthetic Data", "Reinforcement Learning" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/800a939160f39a4a63ad3c3199f90eb6183f3212.pdf" }, "presentation": null, "primary_area": { "value": "unsupervised, self-supervised, semi-supervised, and supervised representation learning" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Self-Informed Generative Active Learning" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. What if the model learns inaccurately in a complex environment?\n\n2. Can you use the normalized score for the experimental results?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. This paper is written well. The method is esay to follow.\n2. This work is evaluted on various domains." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper investigates the development of agents capable of addressing long-horizon tasks through offline model-based reinforcement learning (RL). While current RL methods excel at learning individual skills, they struggle with integrating these skills to accomplish extended tasks due to the mismatch between the termination of one skill and the initiation of another, resulting in distribution shifts. The authors propose an offline approach to skill stitching, leveraging aggregated datasets from various skills to train a dynamics model that can generalize across different skills. This model, along with an ensemble of offline dynamics models and value functions, is used to stitch adjacent skills through model predictive control (MPC). To address the overestimation issues common in offline model learning, a conservative method is introduced to penalize uncertainty in model and value predictions. The study's experimental results demonstrate the effectiveness of this approach over baseline methods in offline settings across multiple benchmarks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The originality of this work is quietly limited. The idea of stitching skills based on value functions is not new; many papers have proposed similar approaches. For example, PEX [1].\n2. A large number of baseline algorithms are missing. For example, OPAL [2] and LPD [3].\n\n\n[1] Zhang, Haichao, We Xu, and Haonan Yu. \"Policy expansion for bridging offline-to-online reinforcement learning.\" arXiv preprint arXiv:2302.00935 (2023).\n\n[2] Ajay, A., Kumar, A., Agrawal, P., Levine, S., & Nachum, O. (2020). Opal: Offline primitive discovery for accelerating offline reinforcement learning. arXiv preprint arXiv:2010.13611.\n\n[3] Yang, Y., Hu, H., Li, W., Li, S., Yang, J., Zhao, Q., & Zhang, C. (2023, June). Flow to control: Offline reinforcement learning with lossless primitive discovery. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 37, No. 9, pp. 10843-10851)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "- I wonder if the value function properly evaluates states that have not been visited (during stitching). As the value functions for each skill are learned distinctly, how can the value evaluation in the stitched space be accurate and reliable?\n- How might the proposed method be adapted to handle low-coverage offline datasets? \n- I wonder if the authors considered any techniques to reduce the computational burden of MPC in continuous or stochastic environment?\n- What potential strategies could be considered for improving generalization or adaptability to dynamic environments within the constraints of offline learning?\n\n- Minor Typos:\n\nline 97: over-estimate → overestimate, to match the usage elsewhere in the paper.\n\nline 215: continous actions space → continuous action space\n\nline 257: T(\\cdot|s_t,a_t) → T_{\\phi}(\\codt|s_t,a_t}" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The proposed offline skill stitching method is straightforward yet effective in certain environments with long-horizon tasks, enabling task completion by sequencing learned skills from offline datasets.\n- Skill stitching offers a practical approach in hierarchical reinforcement learning, addressing challenges in learning tasks composed of multiple sub-tasks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work explores a model-based approach for offline learning of skills and their sequential stitching using only individual skill datasets, without relying on online interactions with the environment. Unlike existing skill stitching techniques based on online reinforcement learning, this approach utilizes offline data to decompose long-horizon tasks into manageable skills that can be executed sequentially. The focus is on training a dynamics model with aggregated skill datasets, enabling effective model-based planning and incorporating conservative optimization objectives to ensure robust transitions between skills during planning." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Lack of novelty: The proposed skill stitching method of evaluating states for stitching using the value function is not novel; it is a fundamental approach used in existing offline RL for trajectory stitching [1, 2]. A comparison with these existing offline trajectory stitching methods is required. \n\n[1] Stitching Sub-trajectories with Conditional Diffusion Model for Goal-\nConditioned Offine RL (AAAI 2024)\n\n[2] Model-based Trajectory Stitching for Improved Offline Reinforcement\nLearning (NeurIPS 2023)\n\nThe below work also uses model-based rollouts (planning) for skill-based task planning in offline settings, similar to the proposed method.\n\n[3] Offline Policy Learning via Skill-step Abstraction for Long-horizon Goal-Conditioned Tasks (IJCAI 2024)\n\n-\tThe proposed method using MPC operates by sampling possible actions and evaluating the value of the resulting states. For continuous action spaces, it requires extensive sampling and evaluation to determine the best outcome. Furthermore, in environments with stochasticity, the MPC optimization can be required at each attempt, leading to significant inefficiencies in time complexity.\n\n- The performance gain in the Kitchen appears minimal, raising questions about whether the proposed method is effective in continuous action space settings. In the Maze Runner, the discrete action space makes the MPC method feasible. However, in complex continuous tasks like the Kitchen task, the value function evaluation may be unreliable, requiring MPC to extensively search the possible action space, which may explain the minimal performance gain observed.\n\n- The method may not generalize well across diverse environments, especially those with dynamic or unpredictable conditions, as it relies solely on offline data without any consideration on real-time adaptability.\n\n- The approach's effectiveness is highly dependent on the diversity of the offline datasets, as the method relies on the learned dynamics model on the aggregated offline datasets." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. For the maze experiments, can you compare to offline goal conditioned RL for example goal-conditioned IQL?\n\n2. For the MF-stitching baseline, do you train the model-free stitching policy for each two adjacent skills?\n\n3. How does the method perform for each skills permutation? Is it better under some permutations and worse in others?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The introduction of the offline skill stitching problem is important for real-world applications\n\n2. The idea of using a model and planning to stitch the skills is interesting and seems a good direction for further research.\n\n3. The method results on the maze are strong compared to the baselines.\n\n4. The results are better than baselines in general." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces an algorithm for skill stitching from offline data,the algorithm has two phases, an offline training phase where each skill is extracted from an offline data that contains trajectories representing the skill. And a test phase, where the dynamics model is used for MPC-based skill stitching guided by the value function. The experiments demonstrate the performance of the method in comparison with some baselines; the ablations show that the quality of the data can have a significant effect on the performance of the skill changing as well as the diversity of transitions in the training distribution." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The assumption of the availability of a dataset for each skill is a strong assumption, is there a way to relax it? For example learning diverse skills from one offline dataset? Is this possible and is there any related work that focus on this problem?\n\n2. Training each skill separately via offline RL seems expensive and time-consuming.\n\n3. For some hyperparameters it is not clear to me they have been chosen, for example the maximum steps of skill execution seems very problem dependent.\n\n4. The method does not seem effective on more complicated tasks (for example in table 2 the method fails in accomplishing more than one skill regardless of the number of skills in the task), but it is still better than the baselines." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024offline,\ntitle={Offline Model-Based Skill Stitching},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=0YxvqG9SsJ},\nnote={under review}\n}" }, "abstract": { "value": "We study building agents capable of solving long-horizon tasks using offline model-based reinforcement learning (RL). Existing RL methods effectively learn individual skills. However, seamlessly combining these skills to tackle long-horizon tasks presents a significant challenge, as the termination state of one skill may be unsuitable for initiating the next skill, leading to cumulative distribution shifts. Previous works have studied skill stitching through online RL, which is time-consuming and raises safety concerns when learning in the real world. In this work, we propose a fully offline approach to learn skill stitching. Given that the aggregated datasets from all skills provide diverse and exploratory data, which likely includes the necessary transitions for stitching skills, we train a dynamics model designed to generalize across skills to facilitate this process. Our method employs model predictive control (MPC) to stitch adjacent skills, using an ensemble of offline dynamics models and value functions. To mitigate overestimation issues inherent in models learned offline, we introduce a conservative approach that penalizes the uncertainty in model and value predictions. Our experimental results across various benchmarks validate the effectiveness of our approach in comparison to baseline methods under offline settings." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Skill stitching", "Offline reinforcement learning", "Model-based planning" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/4da014a566a23e479bb6c134d88301e9dea08e43.pdf" }, "presentation": null, "primary_area": { "value": "reinforcement learning" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Offline Model-Based Skill Stitching" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "see questions above," }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "Uses LSH to approximate attention computation for eviction (if you compare to h2o / scissorhands)" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The idea is to reduce KV Cache by evicting and permanently dropping tokens at each position in the query. The heuristic used is to evict the lowest attention scored keys ( which is essentially similar to H2O / Scissorhands which preserve top attention scored keys). The difference is to use LSH to do a approximate score ranking to avoid SoftMax for exact computation." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Novelty: The novelty is limited.\n- H2O / Scissorhands are known to not perform well on longbenchmark. Can we see some results on longbenchmark like passage retrieval datasets ?\n- Missing baselines --only baseline used is L2 norm. \n- Limited evaluation. can we get more results on longbenchmark at different budgets with standard baselines." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Line 52: \"However, this L2 dropout strategy only performs well on\nlong-context retrieval tasks. It is specialized to retain only those tokens with the highest attention\" -- be more specific. Why is this?\n\nLine 57: \"wide variety of tasks?\" -- how do you define this?\n\nLine 145: Formally for our setup, distd(x, y) cos θx,y, here it is more a measure of cosine similarity than distance. Misleading, perhaps?\n\nLine 419: did you mean \"LSH dimension does significantly impact performance\" --> does not?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "Studies an important problem of much significance in todays LLM era. \n\nPresents a simple yet elegant approach\n\nDoes good evaluations on a range of use-cases" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents new methods to accelerate inference of auto-regressive transformers used in most modern-day decoder-based LLM architectures. Indeed, the main drawback of existing systems is the size of the \"KV Cache\" or Key-Value Cache which is used during the attention mechanism. To speed up the attention calculation, most systems have a cache which remembers the keys and values of commonly used tokens, to avoid recomputing it for each token decoding. However ,such a cache, for it to be performant at inference time, must scale quadratically with the sequence length, and linear in number of layers and attention heads. \n\n(Authors: please explain why for the uninformed reader -- this is stated in the intro, but without explanation)\n\nIn this paper, the authors present an LSH based method to evict far-away key tokens. Indeed, suppose we have an LSH which gets a binary encoding of any vector using random hyperplane projection method (SIMHASH). \nThen, we can first pre-process and compute the hamming distance between query token and all key tokens, and evict the farthest one, as this is the one least likely to affect the overall attention soft-max operation.\n\nThey implement this simple scheme and provide a range of quality vs cache size metrics comparing with one other KV-cache called L2-Dropout Cache, which drops the keys based on their magnitudes." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Why is there no timing experiment, since that will be one key benefit of caching.\n\nWhy only restrict to attention-free cache policies and specifically only compare with the L2-dropout baseline?\n\nConceptually, what is the key difference with Reformer? I have not read that paper but you mention in passing that it is using LSH and simhash also. Is which cells to evaluate vs what to evict the only difference between Reformer and your work? If so, worth comparing with Reformer also in plots?\n\nWhat is the rationale of the policy? Why can't a token just evicted become relevant again? I guess is there some language-based \"locality of reference\"?\n\nDo ablation of the hardcoded bits, i.e., you mention you hard-cache the first few and last few tokens. What is the contribution of this to your overall success metrics?\n\nThe eviction policy is not clearly understandable in how it aggregates the hamming distances over time steps. Is it only based on the most recent time step, or some more complex rule?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "- Could you provide a plot showing the distortion error introduced by LSH compression across different levels of compression? Specifically, how does the approximation quality change as more tokens are evicted or as the quantization parameters are adjusted?\n\n- Given that LSH-E’s efficiency largely depends on its CUDA implementation, can you elaborate on any specific optimizations made within the CUDA code?\n\n- Could you clarify how LSH-E handles multi-head attention? Specifically, is each head processed separately with its own LSH compression, or is there a shared mechanism across heads?" }, "rating": { "value": 1 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The use of theoretical approaches such as SimHash, a highly efficient hashing method, is a valuable aspect of this work, contributing to both the effectiveness and scalability of the proposed method." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces LSH-E, an algorithm for compressing the key-value (KV) cache in large language models (LLMs) using locality-sensitive hashing (LSH). Despite the availability of prior work—including KDEformer, Hyperattention, SubGen, and QJL—that similarly utilizes LSH for efficient attention and memory management, these related efforts are not cited here. LSH-E leverages Hamming distance calculations in a binary space following a Quantized Johnson-Lindenstrauss (JL) transform (SimHash) to identify and evict tokens with low relevance to the current query, resulting in memory savings. This pre-attention approach provides a lightweight, GPU-efficient solution for long-context tasks, although its effectiveness ultimately depends on the algorithm’s CUDA implementation efficiency." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The term \"novel\" should not be used for LSH in this context, as it is not a new approach and has appeared in prior work. Specifically, the methods used in KDEformer, Hyperattention, QJL, and SubGen demonstrate significant overlap, yet these works are not cited here, despite their relevance.\n\n- The experimental setup lacks comprehensiveness; comparisons with alternative methods like H2O, SubGen, and other established baselines should be included to provide a more robust evaluation.\n\n- The datasets used in the experiments are not sufficiently large for evaluating performance in long-context scenarios. Given that these methods target long-sequence processing, experiments should ideally use token sizes over 50,000. LongBench or other large-scale datasets would be more appropriate for a thorough evaluation.\n\n- Additionally, runtime metrics should be reported to assess the efficiency of token generation and to substantiate the computational benefits claimed in the paper.\n\nKDEformer : https://proceedings.mlr.press/v202/zandieh23a.html\nHyperAttention : https://arxiv.org/abs/2310.05869\nSubGen : https://arxiv.org/abs/2402.06082\nQJL : https://arxiv.org/abs/2406.03482" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "See weakness." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "Strong Points\n----\nS1. The problem of the paper is well-motivated. \n\nS2. The proposed algorithm is simple and clear with illustrative example.\n\nS3. The proposed method outperforms the baseline L2." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a method that uses LSH to perform kv cache eviction. The provided experiments show that the proposed method outperforms the baseline." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Weak Points\n----\nW1. Important related studies and baselines are missing:\nSinghania, P., Singh, S., He, S., Feizi, S., & Bhatele, A. (2024). Loki: Low-Rank Keys for Efficient Sparse Attention. arXiv preprint arXiv:2406.02542.\nTang, J., Zhao, Y., Zhu, K., Xiao, G., Kasikci, B., & Han, S. (2024). Quest: Query-Aware Sparsity for Efficient Long-Context LLM Inference. arXiv preprint arXiv:2406.10774.\n\nW2. The key measures of the targeted task should be have more accurate inference with lower memory footprint and latency. I do not agree with the methodology of not comparing with other \"non attention-free\" methods.\n\nW3. The presentation of experiments need to be improved: Lack of discussions and intuitions in the experiment analysis. For example, why does LSH-E outperform Full in Figure 4a; why does LSH-E become worse than L2 after 50% cache budget in Figure 4b? We have many subsubsections in the experiments, but most contents in those are barely text illustration of the figure and result while no discussion of why we would have those results.\n\nW4. The execution time of the proposed system is missing.\n\nW5. The discussion of the error introduced by the LSH is not included. I wonder what if we use cosine similarity to evict the cache instead of LSH, how will be the accuracy, latency, and memory usage?\n\nW6. In the supplementary materials, we see more experiments with more baselines that are better than L2. I wonder the reason why the authors do not include them.\n\n\nPresentation\n----\nP1. Line 180 \"heavy hitters' -> ``heavy hitters''\nP2. The axis captions of the figures are too small to be seen." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Besides the problems mentioned in Weakness,\n1 Does this method work well with quantization (KIVI, AWQ)?\n2 How long does LSH-E increase first token latency?\n\nThese two questions can be left for future work." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "This paper applies novel LSH methods to KV cache problems. The motivations and reasons why LSH can produce a good performance are well discussed. Besides this, a static compression rate of 30% - 70% is also helpful for many LLM serving systems, given the accuracy is preserved." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a KV cache compression method based on LSH and shows that LSH-E can achieve good downstream performance on various downstream tasks with a 30%- 70% compression ratio." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. There is no comparison with other static KV compression baselines, including H2O, streamingLLM, and SnapKV. If this problem is solved, I will raise my score.\n2. Only the memory compression ratio is shown. I will ask for the wall clock speedups (latency or throughput)." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We develop a token eviction strategy that uses locality-sensitive hashing to locate low-attention tokens without computing attention." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024lsh,\ntitle={{LSH} Tells You What To Discard: An Adaptive Locality-Sensitive Strategy for {KV} Cache Compression},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=0ZcQhdyI3n},\nnote={under review}\n}" }, "abstract": { "value": "Transformer-based large language models (LLMs) use the key-value (KV) cache to significantly accelerate inference by storing the key and value embeddings of past tokens. However, this cache consumes significant GPU memory. In this work, we introduce LSH-E, an algorithm that uses locality-sensitive hashing (LSH) to compress the KV cache. LSH-E quickly locates tokens in the cache that are cosine dissimilar to the current query token. This is achieved by computing the Hamming distance between binarized Gaussian projections of the current token query and cached token keys, with a projection length much smaller than the embedding dimension. We maintain a lightweight binary structure in GPU memory to facilitate these calculations. Unlike existing compression strategies that compute attention to determine token retention, LSH-E makes these decisions pre-attention, thereby reducing computational costs. Additionally, LSH-E is dynamic -- at every decoding step, the key and value of the current token replace the embeddings of a token expected to produce the lowest attention score. We demonstrate that LSH-E can compress the KV cache by 30\\%-70\\% while maintaining high performance across reasoning, multiple-choice, and long-context retrieval tasks." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "kv cache", "locality-sensitive hashing", "compression" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/b53389ae75f81b08c5d6441011b7d8c70db2349c.pdf" }, "presentation": null, "primary_area": { "value": "foundation or frontier models, including LLMs" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/5aae791fb7a52587e5f545eb3922e3a5f4031416.zip" }, "title": { "value": "LSH Tells You What To Discard: An Adaptive Locality-Sensitive Strategy for KV Cache Compression" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. The performance and visual results of DefGS and GVFi appear very similar. Could the authors specify scenarios where the translation-rotation dynamics module offers clear advantages?\n2. Could quantitative results for object segmentation be provided, and how does GVFi compare to models that rely on human annotations for this task?\n3. Could the authors highlight the novelty compare to DefGS?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. It is novel to represent the 3D points as particles, which is a well-established concept in robotics. This representation could open up further research topics to improve dynamics modeling.\n2. This model does not rely on human annotations for motion estimation. It can autonomously group meaningful objects based on motion patterns without requiring any labeled data.\n3. The authors provide both quantitative and qualitative results across multiple datasets, demonstrating GVFi’s improvements in both interpolation and extrapolation tasks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces GVFi, a novel approach for modeling 3D scene geometry, appearance, and dynamics from multi-view images without the need for human annotations, such as bounding boxes or segmentations. The authors highlight that previous 3D Gaussian Splatting models struggled to capture the underlying motion physics of dynamic scenes. In contrast, GVFi treats 3D points as particles in space, each with a learnable size and orientation, enabling the model to learn particle rotation and translation to represent a dynamic system effectively. Experimental results on three diverse datasets show that GVFi significantly outperforms prior 3D Gaussian Splatting models on both interpolation and extrapolation tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. This model builds upon DefGS (Yang et al., 2024), with its main contribution being the translation-rotation dynamics system module. However, the novelty of this addition may be somewhat limited.\n2. The performance of DefGS (Yang et al., 2024) and GVFi is quite similar, and there appears to be no significant visual difference between the outputs of the two models. Could the authors clarify specific scenarios where the translation-rotation dynamics system module leads to performance improvements?\n3. There are no quantitative results for object segmentation. Would it be possible to evaluate this and compare it to models that rely on human annotations?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please refer to weaknesses above." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "[+] The paper is well-organized.\n\n[+] The proposed methodology of predicting translation rotation dynamics is straight-forward and well-presented.\n\n[+] The emerged behavior of rigid parts through motion clustering is interesting and show be highlighted further.\n\n[+] Extensive empirical evaluation on multiple benchmarks demonstrates superior performance, along with proper ablation study and demo video in supplementary." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors extend multi-view dynamical scene modeling by predicting motion physics parameters without additional supervision. Specifically, they directly predict a translation rotation dynamics system for each 3D particle, which gives the model capabilities in future predictions of trajectories and rigid part discovery via clustering. Quantitative and qualitative results show superior performance against prior arts on three existing and one proposed benchmarks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "[-] My main concern about this work is the assumption made (L219) that \"there is no additional force involved after $t=0$.\" Although the author give a justification that \"a rolling ball suddenly exploding is not learnable,\" I am not sure if the scope of the research is sufficiently broad given this constraint:\n- First, while some moveable objects cannot move of their own volition, many dynamical (interesting) objects do have the ability to move on their own (e.g. humans, vehicles, animals, etc). By assuming no additional forces after $t=0$, the formulation assumes the presence of no dynamical objects, which conflicts with some of the qualitative results (whale, skater and van). Are we simply modeling these objects in a time window where no force is applied? It would be great if the authors can clarify on how the assumption impacts the modeling of self-propelled objects.\n- Second, due to the strict assumption made about applied forces, the dynamical scene valid for this method would be rather simple and cannot contain more complex motion with evolving accelerations. The authors should elaborate on the types of motion that can / cannot be handled by GVFi.\n- Finally, since I do not work on this topic, I am not sure how significant is my concern above and I am happy to change my recommendation as I await to read other reviewer’s comments and the author's response to my review." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. Based on the methodology, there seem to be three possible approaches for interpolation rendering: (1) directly using $f_{defo}$ to predict the deformation at the given time $t$, (2) progressively calculating the Gaussian deformation at the given time $t$ from time 0 using the motion parameters predicted by $f_{trd}$, or (3) following the steps described in lines L261-L269. Which approach was used in the experiments? Are the results consistent across these three methods?\n2. For extrapolation rendering according to lines L261-L269, it seems feasible to use either the second or third approach from question 1. Which method was actually used by the authors? If the third approach was used, how does it perform over longer extrapolation periods? Could the authors provide visual results for extrapolations that extend beyond the time span covered in the dataset?\n3. The choice of baseline methods for comparison appears limited. For a comprehensive evaluation, it would be beneficial to compare against state-of-the-art methods in dynamic scene reconstruction, such as 4D-GS[2] and more recent work like E-D3DGS [3], which both have architectures similar to Deformable3DGS but differ in their motion representation. Could the authors verify if the proposed Translation Rotation Dynamics System can be integrated into these methods and whether it would yield similar performance gains?\n4. The authors claim that their framework is a general approach for modeling motion physics in complex dynamic 3D scenes. However, the datasets used, with only 60 frames in total, limit the complexity and extent of motion. Could the authors validate this claim by testing on more challenging synthetic and real-world datasets, such as the ParticleNeRF and PanopticSports datasets, to provide a more comprehensive evaluation of the framework’s effectiveness on complex scenes?\n5. In the ablation study, the authors provide a rationale for their choice of $\\delta t$, which is somewhat reasonable. However, this conclusion is based on results from only one dataset, which may not be sufficient, as each dataset could exhibit different motion characteristics. Could the authors clarify how to select an appropriate $\\delta t$ in practice across diverse datasets?\n6. The experimental details are insufficient, particularly regarding training time, required resources, storage size, and rendering speed. Could the authors provide more comprehensive information on these aspects?\n7. Please ensure that all abbreviations and technical terms are clearly defined, with full explanations and necessary citations. In the related work section, it would be helpful to explicitly clarify the differences from relevant works wherever possible.\n\n[2] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024.\n\n[3] Jeongmin Bae, Seoha Kim, Youngsik Yun, Hahyun Lee, Gun Bang, and Youngjung Uh. Per- gaussian embedding-based deformation for deformable 3d gaussian splatting. In Proceedings of the European Conference on Computer Vision (ECCV), 2024." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. Modeling the motion of Gaussians through a Translation Rotation Dynamics System grounded in classical mechanics, resulting in a concise and conceptually elegant framework with solid mathematical and physical foundations.\n2. Introducing an effective method to train the motion parameters of the Translation Rotation Dynamics System, enabling the accurate estimation of translation and rotation dynamics for each particle in the scene.\n4. By explicitly learning motion parameters under classical mechanics, enabling effective extrapolation to unobserved frames and presenting potential for generation tasks that require plausible future frames in dynamic 3D scenes.\n3. The proposed approach is validated on two tasks, demonstrating superior performance compared to previous methods, highlighting its effectiveness in modeling motion dynamics in 3D scenes." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces GVFi, a framework for modeling the motion physics of complex dynamic 3D scenes using multi-view RGB videos without requiring additional annotations such as object shapes, types, or masks.\nBuilding on Deformable3DGS, GVFi incorporates constraints based on the laws of classical mechanics to guide motion predictions, ensuring that the Gaussian deformation estimated by the MLP aligns more closely with physical principles. By assuming that motion adheres to the laws of classical mechanics and explicitly learning the associated motion parameters, GVFi is capable of performing effective extrapolation rendering, allowing it to predict frames beyond the observed time span. Experimental results show that GVFi significantly outperforms existing methods, particularly excelling in future frame extrapolation tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The contributions of this work are somewhat incremental, as most of the methodological design heavily overlaps with the baseline method, Deformable3DGS [1]. The key difference lies in the incorporation of dynamical principles, primarily to enable extrapolation capabilities rather than introducing fundamentally novel approaches.\n2. The proposed motion modeling framework is overly restrictive, relying on an strong assumption of no external forces, disregarding energy transfer processes, and lacking the ability to handle non-rigid or nonlinear motion. These limitations significantly reduce the model’s applicability to real-world physics.\n3. Due to its reliance on idealized assumptions and limited scope, the model struggles to handle complex, real-world motion dynamics where varied forces, interactions, and non-rigid behaviors are prevalent, limiting its utility for practical applications in diverse environments.\n\n[1] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3D Gaussians for High-Fidelity Monocular Dynamic Scene Reconstruction. CVPR, 2024." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "- How long are each of the datasets scenes? Are they really long enough to meaningfully challenge the assumption of second order expansion?\n- The NVIDIA Dynamic Scene Dataset (Yoon 2020) contains many dynamic scenes in the 2020 paper, but this paper claims \"it consists of two real-world dynamic 3D scenes\", what are those scenes?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- Method at its core is quite simple (this is a good thing)\n - Learning a second order taylor series expansion of the full trajectory\n - The quantitative results seem good, even if only minor improvements in a number of cases" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Paper proposes a method \"GVFi\" tackles the problem of estimating dynamic 3D scenes.\n\nBroadly speaking, GVFi\n - Uses an off the shelf method (3DGS) to compute gaussian splats in a canonical frame\n - Uses an off the shelf method (\"Deformable 3D Gaussians for High-Fidelity Monocular Dynamic Scene Reconstruction\" Yang et al., CVPR 2024) to estimate a deformation field over position, rotation, and scale of each gaussian as a function of time\n - Uses these as inputs to then estimate the 3D gaussian's motion \n\nImportantly, these gaussians are parameterized as rotation around a moving rotation centerpoint, and this centerpoint's motion is described entirely by an initial position, velocity, and acceleration estimate. These estimates are then optimized against the flow field as noisy ground truth and training observation reconstruction losses." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Second order taylor series expansion seems quite limiting for arbitrary motion, or motion over non-trivial time horizons\n - Assuming I am interpreting the paper correctly, experiments seem to be only over short (~1 second) time horizons, which don't seem like they would challenge this assumption\n - Presentation quality is *extremely* poor\n - Core concept is quite simple, but it's heavily obfuscated for no apparent reason. It could be explained in 1 paragraph.\n - Core concepts seem poorly motivated; physics priors are common, but why only a second order expansion? Is this really a reasonable assumption in practice? There needs to be more motivation to this choice and more careful analysis of its limitations\n - Figure 1 and 2 are almost the same thing but not very informative. A better figure would be demonstrating the taylor series expansion of a single gaussian's trajectory\n - The math in section 3 does not feel like it was put there to be informative, but instead to intimidate the reader; after climbing through the notation its basically just saying to compose offsets together to estimate motion. If the authors feel this notational exercise is needed (don't think it is), it should go in the appendix and the main paper should have far more explanatory figures.\n - Ablations do not seem to address the core contribution, which is the assumption of the second order expansion --- what if you only do a first order expansion? Can you attempt to extend this to third order? They briefly mention replacing it with an MLP, but minimal details are provided.\n\nI'm of the opinion that the paper has a neat idea but its presentation needs to be dramatically overhauled --- its assumptions need to be clearly stated and examined as reasonable or not, and it needs to have experiments where the method is pushed. Looking at the qualitative results, these datasets are very simple partwise rigid motion and the taylor series expansion is a nice trick to force smooth non-shattering motion, but it comes at the cost of generality --- nowhere does this seem to be addressed, considering the sometimes marginal performance improvements over far more flexible prior methods.\n\nNit:\n\"Cononical\" -> Canonical misspelling is rampant" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024gvfi,\ntitle={{GVF}i: Learning 3D Gaussian Velocity Fields from Dynamic Videos},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=0Zot73kfLB},\nnote={under review}\n}" }, "abstract": { "value": "In this paper, we aim to model 3D scene geometry, appearance, and physical information just from dynamic multi-view videos in the absence of any human labels. By leveraging physics-informed losses as soft constraints or integrating simple physics models into neural networks, existing works often fail to learn complex motion physics, or doing so requires additional labels such as object types or masks. In this paper, we propose a general framework named GVFi to model the motion physics of complex dynamic 3D scenes. The key novelty of our approach is that, by formulating each 3D point as a rigid particle with size and orientation in space, we choose to directly learn a translation rotation dynamics system for each particle, explicitly estimating a complete set of physical parameters to govern the particle's motion over time. Extensive experiments on three existing dynamic datasets and a newly created challenging dataset demonstrate the extraordinary performance of our method over baselines in the task of future frame extrapolation. A nice property of our framework is that multiple objects or parts can be easily segmented just by clustering the learned physical parameters." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Dynamic Reconstruction", "Physics", "Motion Extrapolation" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/60dd1d0ceeb5487f803d2d90706d75c9f91fa9bb.pdf" }, "presentation": null, "primary_area": { "value": "applications to computer vision, audio, language, and other modalities" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/ba93674e0fa0a49347edf567ec9996ff62c59526.zip" }, "title": { "value": "GVFi: Learning 3D Gaussian Velocity Fields from Dynamic Videos" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 2 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please refer to weaknesses." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper is methodologically sound, with a well-defined approach supported by both theoretical and empirical analyses. The experimental setup is robust, including multiple real-world datasets, and the model's performance is compared against established baselines to highlight its predictive strength. \n\nThe paper is well-structured and provides comprehensive explanations of its key components, including the latent utility functions, mixture-of-experts model, and gating function. Diagrams and formulas aid in clarifying complex concepts, making the model's framework accessible for readers. \n\nThis framework contributes significantly to spatial-temporal modeling, especially in domains where human decision-making drives event occurrences. By enabling a nuanced understanding of preference-driven behavior and offering predictive power, the model has applications in fields like criminology, urban planning, and shared mobility systems." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a new spatial-temporal counting process model that integrates choice theory and social intelligence to capture human decision-driven event occurrences, such as crime rates and bike-sharing usage. The core idea is to use latent utility functions to represent diverse decision-making factors and to apply a mixture-of-experts model with a sparse gating function for adaptive selection. The model aims to reveal underlying patterns in counting processes, providing both predictive power and interpretability." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The use of mixture-of-experts and the sparse selection mechanism may raise concerns regarding computational scalability when applied to large-scale, high-dimensional spatial-temporal data. While the model performs well on mid-sized datasets, it is unclear if the sparse gating function and multiple experts could handle significantly larger spatial grids or finer temporal resolutions without substantial computational costs. A discussion on computational efficiency or optimization strategies, such as parallelization, would strengthen the model’s applicability to broader scenarios." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "N/A" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please answer the questions corresponding to the weaknesses mentioned above.\n1. Why use these small datasets for evaluation? What about the actual value of the proposed method when applied to large-scale datasets?\n2. How do you explain the performance improvement of the MoE module and the relation between it and the overall performance improvement?\n3. How about the performance improvement when we have fine-grained spatial grids?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The studied forecasting problem of spatio-temporal events is very important, interesting, and of high value in the real world. \n2. The presentation is overall good, and the organization makes the paper easy to read and comprehend.\n3. The authors select two representative metrics, aRMSE and MAPE, on which the proposed method achieves the best performance among all these models." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper is about the prediction problem for spatial-temporal event data generated by humans. The authors introduced a framework integrating choice theory with social intelligence to model and analyze counting processes. The authors further conducted experiments on several real-world spatio-temporal datasets, and empirical evaluation of crime and bike-sharing datasets demonstrated that the proposed model could achieve the best performance." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The datasets are small, leading to convincing results and conclusions. Although the authors have considered three datasets, NYC Crime, Chicago Crime, and Shanghai Mobike, for evaluation, the scales of these datasets are quite limited. There are only less than 1000 events on the first two datasets, which makes us wonder whether the proposed method can be used in real-world applications where the dataset may be very huge.\n2. The technical contribution of the proposed method is questionable. The proposed method introduces a strategy of MoE, which is widely used in model ensembling and limits the contribution of the whole framework. In other words, it is very likely to improve performance by adding the MoE module. In short, the proposed solution is a bit straightforward.\n3. Figure 2, Figure 3, and Figure 4 require improvement. Observing some informative and insightful conclusions from these figures is very hard since the grids are coarse-grained." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. How do hyperparameter changes, such as learning rate, regularization parameters, and the number of mixture components, affect the model's performance?\n2. In what ways can the model be tested on a variety of datasets with different spatial and temporal characteristics to assess its generalizability?\n3. How can cross-validation and out-of-sample testing be conducted to ensure the model's stability and consistency?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. **Innovative Approach**: The paper introduces an innovative framework that integrates choice theory with social intelligence to model spatial-temporal counting processes. This approach addresses the complex decision-making processes and social factors influencing human-generated event data, such as crime occurrences and bike-sharing activities.\n2. **Interpretable Insights**: The model provides interpretable insights by uncovering latent human preference patterns through utility functions. This feature helps in understanding the underlying mechanisms driving the observed event counts, which is valuable for both academic and practical purposes.\n3. **Predictive Performance**: Empirical evaluations using crime and bike-sharing datasets show that the proposed model achieves good predictive accuracy compared to existing methods. The results indicate that the model can effectively predict event patterns and offer useful insights.\n4. **Theoretical Foundation**: The paper derives a generalization bound that is independent of the number of latent classes, providing a theoretical foundation for the model's robustness and reliability. This theoretical contribution adds to the academic value of the work.\n5. **Practical Flexibility**: The model demonstrates flexibility in handling different types of spatial-temporal data and can incorporate external interventions, making it adaptable to various real-world scenarios." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a novel framework that integrates choice theory with social intelligence to model spatial-temporal counting processes, such as crime occurrences and bike-sharing activities. By capturing latent human preference patterns through utility functions, the model aims to provide deeper insights into the mechanisms driving these events. Empirical evaluations using crime and bike-sharing datasets show that the proposed model offers high predictive accuracy and interpretability compared to existing methods, though potential limitations and future research directions are not extensively discussed." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. **Interpretability Validation**: While the model emphasizes interpretability, this claim is not fully supported with detailed case studies or qualitative analyses. More concrete examples and validation are needed to ensure that the insights provided are actionable and meaningful. Without such validation, the interpretability aspect, though highlighted as a strength, remains somewhat abstract and less convincing.\n2. **Computational Efficiency**: The paper does not extensively address the computational efficiency of the model. Practical applications often involve large-scale datasets, and understanding the model's scalability and resource requirements is crucial. Without this information, it is challenging to determine the feasibility of deploying the model in real-world settings, which could limit its practical utility.\n3. **Future Research Directions**:\nThe paper does not clearly outline future research directions or potential extensions of the model. Discussing these aspects would provide a clearer path for advancing the field and addressing current limitations. Identifying open questions and suggesting avenues for further investigation would enhance the paper's contribution and encourage ongoing research in this area." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Please refer to W3 to W6." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "S1. The experiments are conducted with three real-datasets.\nS2. The writing is fluent and easy to understand." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper aims at including human decision processes and social influence to observe criminal event counts. The proposed model is ambitious to include multiple human decision-making aspects, but the details of formulation and examination are missing. The experimental setup needs further reference to show its practicality." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "W1. Overall, the major concerns are that the paper may not be self-contained and appears disconnected. First, although Fig. 1 visualizes the structure of the proposed model, most details explaining each part are not presented. For instance, what are the differences between spatial and position info? Which model does each expert use? Second, although the abstract and introduction state that social norms, environmental cues, and various other factors are considered, there is no corresponding formulation in Section 3. Finally, the experimental results do not validate these claims either. It is suggested to connect the claims with detailed descriptions in the methods and experiment sections.\n\nW2. Please include up-to-date related works in top journals [1][2][3]. Moreover, half of the comparative baselines in the experiment section were published more than 10 years ago, which may be too outdated for fair comparisons. It is suggested to compare with newer methods instead.\n[1] Weichao Liang, Zhiang Wu, Zhe Li, Yong Ge: CrimeTensor: Fine-Scale Crime Prediction via Tensor Learning with Spatiotemporal Consistency. ACM Trans. Intell. Syst. Technol. 13(2): 33:1-33:24 (2022)\n[2] Shuai Zhao, Ruiqiang Liu, Bo Cheng, Daxing Zhao: Classification-Labeled Continuousization and Multi-Domain Spatio-Temporal Fusion for Fine-Grained Urban Crime Prediction. IEEE Trans. Knowl. Data Eng. 35(7): 6725-6738 (2023)\n[3] Weichao Liang, Jie Cao, Lei Chen, Youquan Wang, Jia Wu, Amin Beheshti, Jiangnan Tang: Crime Prediction With Missing Data Via Spatiotemporal Regularized Tensor Decomposition. IEEE Trans. Big Data 9(5): 1392-1407 (2023)\n\nW3. The definitions of matrices A and B on line 227, page 5, and the purpose of formulating them are unclear. Specifically, what do the two matrices embed, respectively? Additionally, right before introducing these matrices, the model already includes positional, spatial, temporal, and feature embeddings. An alternative approach might be to directly feed these four embeddings to the experts, rather than combining them with the two matrices to avoid additional computational overhead. This raises questions about the necessity, purpose, and benefit of the intermediate matrix decomposition-based embedding method compared to a straightforward alternative.\n\nW4. Please clarify the “ranking” concept in the gating function, starting from line 251 on page 5. Equations 7, 8, and the loss function at line 274 resemble a cross-entropy formulation, which is a classification-based metric rather than a ranking one. Additionally, I am uncertain whether ranking is appropriate in this scenario. Specifically, while predicting the time and place of a crime, a top-1 ranking for occurrence may not directly indicate that a crime is happening, as the probability could still be low. Therefore, relying on ranking rather than probability prediction may lead to false alarms and overreactions.\n\nW5. The practicality of the experimental setup is questionable. In the New York Crime and Chicago Crime datasets, each city is divided into 100 areas, and daytime is segmented into 4 time slots. However, it is unclear how large each area is after division. Is there evidence or a reference supporting that the 100-block granularity is beneficial for real-world law enforcement? Similarly, dividing daytime into four 6-hour slots may not be sufficiently granular. Is there a reference justifying this setup? Furthermore, it would be interesting to see the model’s performance at finer granularities, with smaller areas and shorter time slots.\n\nW6. The experimental results may not fully examine the authors' claims. While modeling the “human decision process” is a key focus, it is unclear how this is tested in the experiments. Are there specific sequential criminal events in the datasets? If so, does the proposed method successfully retrieve these sequences? How does the model demonstrate that its improvements are due to modeling the human decision process? Otherwise, if each crime is independent, how are the datasets suitable for examining causal relationships? In this context, could simple statistics identify criminal hotspots at specific time slots to yield similar results to those in Fig. 2? It is recommended to elaborate further on human decision modeling in the experiments." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024preferencedriven,\ntitle={Preference-Driven Spatial-Temporal Counting Process Models},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=0a7TRHhhcS},\nnote={under review}\n}" }, "abstract": { "value": "Traditional spatial-temporal models often overlook the complex decision-making processes and social factors that shape spatial-temporal event data generated by humans. This paper introduces a novel framework that integrates choice theory with social intelligence to model and analyze counting processes, such as crime occurrences or bike-sharing activity, where the observed discrete events result from individual decisions influenced by social dynamics. \nOur approach aims to uncover latent human preference patterns, represented by utility functions, to capture the diverse decision-making factors within a population that result in the observed event counts. These latent factors help explain how choices—such as where and when to commit a crime—are shaped by personal preferences, environmental conditions, and social influences. By modeling the aggregate outcomes of these individual choices, we can better understand and predict patterns in counting processes. The proposed model adopts a preference-driven approach to counting data, providing interpretable insights at a detailed level. It also enables in-depth analysis of how external interventions, like law enforcement actions or policy changes, influence individual decisions and how these effects spread through the system. Empirical evaluation of crime and bike-sharing datasets demonstrates our model's ability to offer clear insights and achieve high predictive accuracy." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "choice model", "spatial-temporal counting process model" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/0eebdce162a74af0ed30ff858758c2f72e00851c.pdf" }, "presentation": null, "primary_area": { "value": "interpretability and explainable AI" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Preference-Driven Spatial-Temporal Counting Process Models" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "besides the above, I am curious why the perturbation radii are set the way they are, ie. inner rho twice the outer rho? Was it grid -searched? is there some intuition behind this setting?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The motivation to combine the element of sharpness-minimization in meta-learning for better generalization makes sense. This is the operationalized well in the form of an algorithm that is shown to perform slightly better with the baselines. \n\n- The method seems to be extensively tested in supervised learning setups, meta-learning scenarios, as well as those with label noise." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work aims to combine Sharpness-Aware Minimization with Model-Agnostic Meta-Learning, by having worst-case robustified versions of the loss in both the inner and outer loop of meta-learning. This is then tested in the usual supervised learning setups in vision and some meta-learning benchmarks, where the method is shown to outperform the baselines." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Difference wrt Abbas et al. 2022: When compared with this prior work, it is unclear what is the novelty here. The authors mention this paper, but don't bother to explain the similarities or differences. The method here looks **eerily similar** to the two-year old prior work, which is arguably better written and presented and a lot more richer. Except for little bits of analysis on congruence between gradients, I can't spot much of a methodological difference. \n\n- Supervised learning experiments: In departure to their motivation, the authors start presenting results on supervised learning. I understand that this can be simulated in the meta-learning setup as well, but it comes across as confusing. Then their method involves 4 gradient computations per step, while SGD and SAM will involve 1 and 2 gradient computations respectively. So for a fairer comparison, the authors should have reported results with letting the baselines have more compute. Thus, given the excessive runtime, the method does not seem worth the effort of obtaining marginal gains. \n\n- Ablation studies on the relevance for SAM in inner/outer stages of meta-learning would have been insightful: Which part benefits more from SAM? Can the authors run an ablation study?\n\n- Momentum hyperparameter and it's ablation: The Table 8 would suggest that having no momentum results in better performance, but it is bizarre that the authors continue to keep using a momentum in all their results, despite of that. Especially when the improvements they report are not infrequently of the similar range." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "- Can authors state exactly what contributions to claim from Theorems 1 and 2?\n- (on the first Imagenet experiments) the top-1/top-5 accuracies seem quite low, why is that the case? can the authors also provide Resnet50 results? how many runs are these results? can authors provide standard errors?\n- It is unclear the exact difference between two versions of Agnostic-SAM (Table 2 and Table 5); if it means using different base SAM (i.e., SAM or ASAM), it appears that Agnostic-SAM (with ASAM) often underperforms Agnostic-SAM (with SAM), why is it the case?\n- How did author come up with the rules to set perturbation bounds originally?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- SAM and MAML are both found to be effective for enhancing generalization performance, and that the paper is attempting to explore the intersection of these is encouraging.\n- The paper follows a standard procedure to evaluate the proposed method (Agnostic-SAM) and shows its effectiveness in experiments." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper combines MAML into SAM, proposing a new optimization scheme to improve generalization performance. The paper provides a theoretical propositions on generalization bound and gradients alignments, but it is regarded that the paper mainly focuses on verifying its generalization effectiveness numerically measured on some deep learning tasks. The paper also provides additional ablation results to support gradient alignments and momentum." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "There are several concerns on this paper summarized as follows.\n\nMethod\n- The main idea and motivation of this work, as its current form, remain quite random. They are two of many potential ways to improve generalization performance, but without clearly justifying why these two, the paper simply combine the two approaches and end up providing experimental results. This diminishes the technical contributions and novelty.\n- The authors also claim that it is a \"framework\", but with it being the simple combination of SAM and MAML, it has not been rigorously evaluated to be a framework as to whether this can serve as a general scheme so it remains as an initial idea. There have been many advancements since the original SAM and MAML, but the paper only takes a proof-of-concept approach, limiting its potential.\n- This idea requires additional computations (validation set, additional forward-backward, hyperparameter tuning) but it is unclear whether this is worth, in particular, compared to other potential ways to improve generalization.\n\nExperiments\n- The experiments are also a bit bland without being tailored to specifically analyze any aspect of SAM and MAML simply evaluating the final performances, lacking novelty and interesting insights.\n- The proposed scheme is only compared to naive baselines, and it is seen that the improvements are very marginal across many experiments. It is a bit critical in the sense that Agnostic-SAM makes use of the additional validation set and more computations to get validation gradients, which leaves a question that whether Agnostic-SAM is really the best possible choice for generalization." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Please refer to the concerns raised in the weaknesses section above." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The paper provides a comprehensive evaluation of Agnostic-SAM across a wide range of tasks, including image classification, transfer learning, training with label noise, and meta-learning." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces Agnostic-SAM, an optimization method that integrates insights from MAML into SAM. The approach seeks to update the model to a region that not only minimizes sharpness on the training set but also implicitly ensures strong performance on the validation set." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.\tThe motivation for the problem formulation in Equation 3 is not convincingly justified. It would benefit from a clearer explanation of why this specific formulation was chosen and how it directly leads to generalization.\n2.\tThe paper does not sufficiently clarify how the integration of MAML’s insights with the proposed problem formulation and algorithm specifically aids generalization. A deeper theoretical or empirical justification is needed.\n3.\tThe proposed algorithm assumes the existence of a held-out validation set. However, in practice, the training set is used as the validation set, which diverges from the theoretical framework. This discrepancy is particularly problematic in datasets like CIFAR-10 and CIFAR-100, where the training loss converges to zero, a behavior not typically observed with a true validation set.\n4.\tFrom Algorithm 1, it appears that Agnostic-SAM requires double the computational time compared to SAM. In the experiments, SAM is compared by allowing SGD to run for double the iterations for fair comparison [1]. It would be fairer to allow SAM and ASAM to run for twice the iterations of Agnostic-SAM in the experiment.\n\n[1] Foret, P., Kleiner, A., Mobahi, H., and Neyshabur, B. Sharpness-aware minimization for efficiently improving generalization. In International Conference on Learning Representations, 2021" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "According to Section 3.3 the goal is to align the perturbed gradients of the validation and train batches. Why do the authors then report the alignment of the unperturbed gradient of the train batch with the perturbed gradient of the validation batch in Section 5.1?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The proposed method requires an additional hyperparameter, but the authors found a way of setting it consistently throughout their experiments: $ \\rho_{1} = 2 \\rho_{2} $.\n\n- Agnostic-SAM improves over baselines in most cases (even though I have doubts about the setups, see below)\n\n- Combining ideas from MAML and SAM is a creative approach" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors propose Agnostic-SAM, a new variant of the Sharpness-Aware Minimization (SAM) algorithm. Instead of only the batchwise gradient ascent perturbation step of SAM, Agnostic-SAM additionally performs a descent step on a validation batch, before computing the gradient for the final update. The authors motivate their work from a PAC-Bayes bound and report experimental results on image classification tasks (vanilla classification, noisy labels, meta-learning)." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "**Comparison to Baselines**\n\nAt its core, Agnostic-SAM changes the perturbation step of SAM by adding an additional perturbation based on gradients from a separate, smaller data batch, and the authors claim improved generalization performance. However, several methods have proposed adjustments to SAM’s perturbation model with improved generalization performance. Most similar to Agnostic-SAM, [1] adds random perturbations to the gradient-based perturbation, while [2] and [3] perform multi-step perturbations. Many more methods exist ([6,7,8,...]), but none of those appear as baselines in the experiments. The only other standard baselines are in Table 2 (ASAM), and in Table 5 (ASAM and FSAM). In the MAML experiment (Tables 6 and 7) the authors report improved performance over [9]. However, they only show the Sharp-MAML-low version from [9], even though other versions exist. In particular, the Sharp-MAML-both variant from [9] outperforms Agnostic-SAM in three out of the four reported cases. It is unclear why this is not reported or explained. \n\n\n**Training time**\n\nThe proposed method requires two additional forward-backward passes on the validation batch, which leads to increased computational cost compared to SAM (roughly 27\\% wall clock time according to Table 9). While the authors briefly mention this in the conclusion, a more thorough discussion and evaluation is needed, as this affects the fairness of comparisons in the main paper.\n\n\n**Hyperparameters and train settings**\n\nThe authors report some baseline values from the original papers, and others are reproduced with the $\\rho$ values from those papers. For instance, for WRN28-10 on CIFAR100 the SGD number is taken from the SAM paper [5], and the SAM number is reproduced with the same $\\rho$ value, but is lower than the number from the SAM paper (83.0\\% vs 83.5\\% in the SAM paper, which would outperform the reported Agnostic-SAM number). Similar observations hold for ASAM, and CIFAR100. Lower reproduced numbers can be due to different training settings and are not a problem per se if the comparison is fair, but here I have certain doubts because the optimal $\\rho$ value can be sensitive to the training settings and was taken from the reference papers. Further, some choices, like e.g. $\\rho=0.05$ for SAM in ImageNet transfer learning while $\\rho=0.1$ for ImageNet training from scratch look just arbitrary. Some $\\rho$ tuning must have taken place, since the authors even claim that _accuracies tend to decrease when reducing $\\rho$_. Further, it is unclear how exactly the authors came up with the choice $\\rho_{1}=2\\rho_{2}$ and if it is based on the ablation in A2, purely by intuition or additional experiments and tuning. Finally, the scope of the experiments is somewhat limited. In particular, there are no experiments with VisionTransfomers, no experiments on text data, and the only larger-scale experiments (ImageNet) are with fairly weak models (at most ResNet-34 for training from scratch).\n\n\n**Theorem 1**\n\nThe authors present Theorem 1 as a central motivation for their method. However, this theorem is nearly identical to Theorem 1 and its proof in the original SAM paper [5], with minimal modification. As with [5], this theorem would theoretically motivate a version of SAM based on average-case rather than worst-case perturbations. The presented generalization bound implies an average-case sharpness bound, which is only subsequently upper-bounded by a worst-case sharpness bound. This limitation was already present in [5] and has since been highlighted, for example, in [4]. Furthermore, the conclusions from this theorem, i.e. why exactly it would motivate equation (3) and the final algorithm, are not understandable to me. \n\n\n**Clarity**\n\nApart from the disconnect between Theorem 1 and the method, it is not well justified why exactly the alignment of the gradients of the perturbed points from train and validation batches would be beneficial for generalization, especially since in the experiments both batches are from the train set. Overall, the MAML perspective is unclear to me, since in the practical algorithm, train and validation batches are both sampled from the train set, and there is only one task to solve in almost all experiments. Additional confusion arises from unclear terminology (e.g. the notation $\\theta^{*}(\\theta)$ wasn’t introduced, the Taylor expansion in (7) is presented as an exact equality, etc.)\n\n\n\n[1] Yong Liu, Siqi Mai, Minhao Cheng, Xiangning Chen, Cho-Jui Hsieh, & Yang You (2022). Random Sharpness-Aware Minimization. In Advances in Neural Information Processing Systems.\n\n[2] Kim, H., Park, J., Choi, Y., Lee, W., and Lee, J. Exploring the effect of multi-step ascent in sharpness-aware minimization\n\n[3] Goncalo Mordido, Pranshu Malviya, Aristide Baratin, & Sarath Chandar (2024). Lookbehind-SAM: k steps back, 1 step forward. In Forty-first International Conference on Machine Learning.\n[4] Maksym Andriushchenko and Nicolas Flammarion (2022). Towards Understanding Sharpness-Aware Minimization. ICML 2022\n\n[5] Pierre Foret, Ariel Kleiner, Hossein Mobahi, and Behnam Neyshabur. Sharpness-aware minimization for efficiently improving generalization. In ICLR, 2021\n\n[6] Minyoung Kim, Da Li, Shell X Hu, and Timothy Hospedales. Fisher SAM: Information geometry and sharpness aware minimisation. In Kamalika Chaudhuri, Stefanie Jegelka, Le Song, Csaba Szepesvari, Gang Niu, and Sivan Sabato (eds.), Proceedings of the 39th International Conference on Machine Learning\n\n[7] Mi, P.; Shen, L.; Ren, T.; Zhou, Y.; Sun, X.; Ji, R.; and Tao,D. 2022. Make Sharpness-Aware Minimization Stronger: A Sparsified Perturbation Approach\n\n[8] Jiawei Du, Hanshu Yan, Jiashi Feng, Joey Tianyi Zhou, Liangli Zhen, Rick Siow Mong Goh, & Vincent Tan (2022). Efficient Sharpness-aware Minimization for Improved Training of Neural Networks. In International Conference on Learning Representations.\n\n[9] Momin Abbas, Quan Xiao, Lisha Chen, Pin-Yu Chen, and Tianyi Chen. Sharp-maml: Sharpness-aware model-agnostic meta learning" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We introduce a novel optimizer Agnostic-SAM that adapts the core idea of SAM by optimizing the model toward wider local minima using training data, while concurrently maintaining low loss values on validation data." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024agnostic,\ntitle={Agnostic Sharpness-Aware Minimization},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=0aTIvSJ83I},\nnote={under review}\n}" }, "abstract": { "value": "Sharpness-aware minimization (SAM) has been instrumental in improving deep neural network training by minimizing both the training loss and the sharpness of the loss landscape, leading the model into flatter minima that are associated with better generalization properties. In another aspect, Model-Agnostic Meta-Learning (MAML) is a framework designed to improve the adaptability of models. MAML optimizes a set of meta-models that are specifically tailored for quick adaptation to multiple tasks with minimal fine-tuning steps and can generalize well with limited data. In this work, we explore the connection between SAM and MAML in enhancing model generalization. We introduce Agnostic-SAM, a novel approach that combines the principles of both SAM and MAML. Agnostic-SAM adapts the core idea of SAM by optimizing the model toward wider local minima using training data, while concurrently maintaining low loss values on validation data. By doing so, it seeks flatter minima that are not only robust to small perturbations but also less vulnerable to data distributional shift problems. Our experimental results demonstrate that Agnostic-SAM significantly improves generalization over baselines across a range of datasets and under challenging conditions such as noisy labels or data limitation." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "sharpness-aware", "agnostic model", "optimizer", "MAML", "SAM" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/049b2dbe838e5e43b39454e757744ec0a45b5adf.pdf" }, "presentation": null, "primary_area": { "value": "optimization" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Agnostic Sharpness-Aware Minimization" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "None" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please address my concerns in the weakness part, especially the novelty of the proposed method and the theoretical foundations." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The paper is well-written and presents the core ideas in a clear and accessible manner.\n- Using a \"landscape\" to describe the benefits of unconstrained models is particularly novel and insightful." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper investigates unconstrained models for handling data symmetry. The authors demonstrate that by designing a loss function specifically tailored for learning equivariance, unconstrained models can approximate data symmetry by minimizing this equivariant loss. This approach allows the models to efficiently control the level of equivariance while maintaining flexibility." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The study introduces this equivariant loss without providing a strong theoretical foundation for the proposed approach. Also, the proposed method is quite straightforward, and its distinction from data augmentation is unclear. It essentially computes the loss on a larger augmented dataset by sampling transformed data. I suspect this method is already well-known within the community, which limits the novelty of the contribution.\n\n- The experimental comparisons are performed on a limited set of classic models rather than state-of-the-art models, raising concerns about the practical applicability of the method to more advanced techniques." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. To address fundamental concerns, I recommend a more comprehensive background and analysis, alongside broader empirical comparisons. Specifically, the study should include a wider range of architectures that go beyond strictly equivariant and non-equivariant models, particularly for the motion capture task, which is inherently non-equivariant.\n\n2. In practice, if a single randomly sampled group element is used per sample in each training step (as mentioned at the end of Section 3.1), this should be explicitly stated in Section 3.2 where the sampling procedure is discussed and the number of samples M is introduced.\n\n3. The paper lacks heuristic, theoretical and/or empirical justification for the choice of one group element per sample per training epoch sufficient. As a result the particular choice of measure is understudied. Moreover, it remains unclear how the equivariance error is measured in Section 6. How many samples are used for this computation?\n\n4. It's entirely unclear if the difference in the loss landscape is a result of the augmented loss function or a result of the architectural difference. I would strongly recommend performing comparisons with a fixed architecture.\n\n5. The MD17 dataset includes two regression targets: energy and forces. Please clarify in the text which target is being used (likely force regression) and how the results are generated.\n\n6. Near the end of Section 6.2, the statement \"Best performance is observed at an intermediate level of equivariance...\" is confusing. Since the paper modifies the loss function, not the architecture, this needs further explanation for supporting the proposed methodology. Otherwise, the conclusion is simply to not utilize strict equivariant architectures for non-equivariant tasks." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The augmented loss function is generalizable across various architectures.\n\n2. The augmented loss function requires relatively few samples to work effectively making it computationally efficient." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces an augmented loss function that incorporates a measure of average equivariance, aiming to enhance the prediction of approximately equivariant information. The authors validate this approach on both equivariant and non-equivariant tasks, utilizing transformer and graph neural network architectures. Furthermore, they conduct a visual analysis of the loss landscape, comparing two different architectures: Transformer with the augmented loss and GATr without." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The proposed methodology can be described as approximate equivariance, but the paper lacks an adequate background and comparative analysis against existing works on approximate equivariance. This raises concerns about both the novelty and the empirical validation of the approach.\n\n1. Novelty: Augmented loss functions enforcing approximate equivariance have been studied (e.g. [1]) including an average measure (e.g. [2]).\n\n2. Empirical Support: The paper does not benchmark against other methods that address approximate equivariance (e.g., [1]), nor does it consider theoretically grounded approaches to symmetry breaking (e.g., [3], [4]) or simpler strategies like combining SE3Transformer with MLPs.\n\n[1] Kim, Hyunsu, Hyungi Lee, Hongseok Yang, and Juho Lee. \"Regularizing towards Soft Equivariance under Mixed Symmetries.\" Proceedings of the 40th International Conference on Machine Learning, ICML'23, 2023, pp. 686, JMLR.org.\n\n[2] K. Lin, B. Huang, L. M. Collins, K. Bradbury and J. M. Malof, \"A simple rotational equivariance loss for generic convolutional segmentation networks: preliminary results,\" IGARSS 2019 - 2019 IEEE International Geoscience and Remote Sensing Symposium, Yokohama, Japan, 2019.\n\n[3] Wang, Rui, Robin Walters, and Tess Smidt. \"Relaxed Octahedral Group Convolution for Learning Symmetry Breaking in 3D Physical Systems.\" NeurIPS 2023 AI for Science Workshop, 2023, https://openreview.net/forum?id=B8EpSHEp9j.\n\n[4] Lawrence, Hannah, Vasco Portilheiro, Yan Zhang, and Sékou-Oumar Kaba. \"Improving Equivariant Networks with Probabilistic Symmetry Breaking.\" ICML 2024 Workshop on Geometry-grounded Representation Learning and Generative Modeling, 2024, https://openreview.net/forum?id=1VlRaXNMWO." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- How will your algorithm fare when there are data constraints? Equivariant models are inherently data efficient, but your algorithm does not seem to be.\n- The loss landscape plots depend on the selected directions - so how can we infer from just two random directions that the loss landscape is better for Transformers or GATr? The optimization paths should affect, and although it is discussed to some extent in limitations, it would be better if there is more discussion on this.\n\nMost of the other questions I had are listed in the Weaknesses section. I will be happy to improve the score if the authors address the questions and weaknesses with supportive evidence during the discussion phase." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The paper is well-motivated and clearly written (particularly the sections on background and methods).\n- The limitation section discusses an important limitation of the interplay between optimization paths and loss landscape.\n- The experiments are conducted in different domains and examine several essential aspects of the algorithm, giving more insights into the method and how levels of equivariance can affect downstream task performance." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper attempts to build equivariance into unconstrained models by posing equivariance as a constrained optimization problem, which can, in turn, also control the level of approximate equivariance in the models. The authors demonstrate results in N-body dynamical systems, motion capture, and molecule dynamics, and they analyze the effect of the level of approximate equivariance on task performance." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- **Related work**:\n - Although the paper uses equivariance as a constrained optimization problem and discusses it in the context of unconstrained models, it misses several crucial relevant works. Discussion of these works would help to place the submission in the literature and give a view of how this work differs from and compares to existing works.\n - Learning equivariance from data [1, 2], approximate/soft equivariance [3, 4, 5, 6], equivariance as a constrained optimization problem [7, 8], and equivariance with unconstrained models [9, 10, 11, 12].\n - Can the authors highlight the differences from Sec 3.1 and Sec. 3.2 of [10]?\n\n- **$\\beta$ and $\\alpha$ as hyperparameters**:\n - The authors suggest that the level or extent of equivariance can be controlled with $\\beta$ and $\\alpha$ - is there a formal way to define this \"level\" of equivariance or is it an intuition tied to the loss itself, i.e., higher $\\frac{\\beta}{\\alpha}$ indicates more equivariant? \n - Next, how would someone know the optimal level of equivariance while using your proposed algorithm - $\\beta$ is not learned, and the results indicate that the optimal $\\beta$ can be identified from the test data results, which is not ideal. Rephrasing this, how do you know how much equivariance is required for the task, and thus what values of $\\alpha$ and $\\beta$ to set?\n\n\n- **Methodology**:\n - How will your algorithm work if group $G$ is unknown?\n - How can your method reasonably approximate equivariance if $G$ is very large and the duration of training is not enough?\n - The highest level of equivariance is when $\\alpha = 0$ and $\\beta=1$. However, this is equivalent to data augmentation, which does not guarantee exact equivariance. Can your algorithm guarantee exact equivariance?\n - While the trends are consistent for both metrics, as reported in the paper, it might be helpful to discuss which metric - Eq. 9 or Eq. 10 is better suited for evaluation. How does Equation 9 work (or make sense) when $f(x)$ is non-scalar?\n - For Motion Capture, if the symmetry constraints are already known, instead of complete SE(3) equivariant baselines, why didn't the authors select appropriate equivariant models that are equivariant to the required SE(3) subgroup or consider symmetry breaking [14, 15]? What $G$ did your algorithm use? If it is the subgroup of SE(3), then it is an unfair comparison.\n\n- **Minor spelling errors**:\n - L156 \"requiring equivariant into\" should be \"requiring equivariance in\"\n - L396 \"it is\" should be \"its\"\n\n\n**References**:\n1. Equivariance Discovery by Learned Parameter-Sharing. Yeh et al., AISTATS 2022.\n2. Learning Equivariances and Partial Equivariances from Data. Romero et al., NeurIPS 2022.\n3. Learning Layer-wise Equivariances Automatically using Gradients. Ouderaa et al., 2023.\n4. Residual Pathway Priors for Soft Equivariance Constraints. Finzi et al., NeurIPS 2021.\n5. Almost Equivariance via Lie Algebra Convolutions. McNeela et al., 2024.\n6. Regularizing Towards Soft Equivariance Under Mixed Symmetries. Kim et al., ICML 2023.\n7. Improved Canonicalization for Model Agnostic Equivariance. Panigrahi et al., 2024.\n8. Structuring Representations Using Group Invariants. Shakerinava et al., NeurIPS 2022.\n9. Equivariance with Learned Canonicalization Functions. Kaba et al., ICML 2023.\n10. Equivariant adaptation of large pretrained models. Mondal et al., NeurIPS 2023.\n11. Equi-Tuning: Group Equivariant Fine-Tuning of Pretrained Models. Basu et al., AAAI 2023.\n12. Learning Probabilistic Symmetrization for Architecture Agnostic Equivariance. Kim et al., NeurIPS 2023.\n13. Steerable Equivariant Representation Learning. Bhardwaj et al., 2023\n14. Symmetry breaking and equivariant neural networks. Kaba et al., NeurIPS NeuReps workshop 2023\n15. Improving Equivariant Networks with Probabilistic Symmetry Breaking. Lawrence et al., ICML GRaM workshop 2024." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "See weekness." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "Relaxing equivariance is a valuable research direction that can break through the constraints on generalization or expressive power caused by strictly equivariant operations." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this paper, the authors analyze the application of unconstrained models in equivariant tasks by conducting a comprehensive analysis of unconstrained models, comparing their performance and computational efficiency against equivariant models. Besided, the authors introduce a novel, simple loss function that enables these models to approximate symmetries, which can be optimized during training." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The primary concern is the authors' motivation. The idea of using group transformations for data augmentation is native, but for many equivariant tasks, it is challenging to obtain a general model through data sampling due to the bias introduced by limited sampling. For instance, for point clouds or molecules, sampling across all angles would expand the dataset by hundreds of times and still struggle to enable the model to effectively learn fine-grained rotation equivariance. I suggest the authors validate their approach on common 3D datasets such as QM9 or ModelNet40.\n\n2. The authors base their introduction in the first three sections on general equivariance, yet the impact of different equivariance groups on algorithms varies. For example, permutation equivariance and translation equivariance can be directly covered by simple operations, making the paper's method inapplicable. The authors should specify which equivariant tasks their method focus on.\n\n3. Relaxing equivariance is a widely discussed topic, and the authors lack relevant citations and analysis [1] [2] [3] [4]. Moreover, the main advantage of unconstrained models is their ability to learn more complex features. It is worth noting that strictly equivariant operations can limit the expressive power of GNNs [5] [6], but unconstrained models may surpass these limitations. Additionally, some tasks are not strictly equivariant, allowing unconstrained models to be applicable. The authors' emphasis on the lower computational complexity of unconstrained operations is incorrect. In the 3D domain, models like torchmd are strictly equivariant yet have low complexity.\n\n [1] Residual pathway priors for soft equivariance constraints, Finzi, et al.\n\n [2] Approximately equivariant networks for imperfectly symmetric dynamics. Wang, et al.\n\n [3] Relaxing equivariance constraints with non-stationary continuous filters. van der Ouderaa, et al.\n\n [4] Learning Partial Equivariances from Data. Romero, et al.\n\n [5] On the Universality of Rotation Equivariant Point Cloud Networks. Nadav Dym, Haggai Maron. \n\n [6] On the Expressive Power of Geometric Graph Neural Networks. Chaitanya K. Joshi, Cristian Bodnar, Simon V. Mathis, Taco Cohen, Pietro Liò.\n\n4. I do not understand how the loss surface in Figure 1 was created and why it demonstrates the advantages of unconstrained models.\n\n5. There are numerous issues with the paper's presentation:\n\n (a) The equations in lines 218, 224, and 227 lack numbering.\n\n (b) In line 215, the definition of G is finite, which is problematic for integrals where the group size can be infinite. Most groups mentioned in the paper are infinite, and I do not understand why the authors restrict groups to being finite in their initial definition.\n\n (c) All the references have formatting issues because none of them specify the source of the papers. For instance, \"Equivariant Graph Hierarchy-Based Neural Networks\" in your paper is accepted in NeurIPS 2022, not arxiv.\n\n (d) Appendix B is incomplete; several titles are clustered together without any explanatory text." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024learning,\ntitle={Learning Symmetries through Loss Landscape},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=0aaaM31hLB},\nnote={under review}\n}" }, "abstract": { "value": "Incorporating equivariance as an inductive bias into deep learning architectures, to take advantage of the data symmetry, has been successful in multiple applications such as chemistry and dynamical systems. The build of equivariance architecture, particularly w.r.t. roto-translations, is crucial for effectively modeling geometric graphs and molecules, where the understanding of 3D structures enhances generalization. However, despite their potential, equivariant models often pose challenges due to their high computational complexity. In this paper, we study the capabilities of unconstrained models (which do not build equivariance into the architecture) and how they generalize compared to equivariant models. We show that unconstrained models can learn approximate symmetries by minimizing additional simple equivariance loss. By formulating equivariance as a new learning objective, we can control the level of approximate equivariance in the model. Our method achieves competitive performance compared to equivariant baselines while being 10x faster at inference and 2.5x at training." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Unconstrained models", "equivariant models", "symmetries." ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/d4b77b8f4bd303b6beaa3c81a10b3e38e7e05e45.pdf" }, "presentation": null, "primary_area": { "value": "learning on graphs and other geometries & topologies" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Learning Symmetries through Loss Landscape" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "- **Why was Vocos selected?** Given that other Vocoders with potentially better performance are available, why was Vocos specifically chosen? Additionally, what was the necessity of switching the decoder from Encodec’s original model?\n\n- **Reason for the Significant Improvement in SIM**: It is understandable that Repetition Aware Sampling could lead to an improvement in WER; however, it is less clear how this would directly impact SIM. Furthermore, why does the subjective evaluation show significant improvement despite relatively poor objective metrics?\n\n- **Lack of Evaluation on Difficult Cases**: The introduction references challenging cases, yet no evaluation related to these cases is provided. Why is this evaluation absent?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The paper is written in a clear and accessible manner, enhancing readability and comprehension. Notably, VALL-E 2 demonstrates superior performance over ground truth in subjective evaluations, achieving higher scores in both CMOS and SMOS on both the LibriSpeech test-clean and VCTK datasets." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "VALL-E 2 is an LM-based TTS model based on VALL-E. It proposes two new methods:\n\n1. **Repetition Aware Sampling**: In this method, during the sampling process, the repetition ratio is calculated based on the number of times a token has been generated. If this value exceeds a threshold, tokens are generated randomly from the original distribution.\\\\\n2. **Grouped Code Modeling**: This method reduces sequence length by grouping adjacent tokens into fixed-size groups.\n\nThanks to these contributions, VALL-E 2 achieves significantly higher performance than the baseline VALL-E, particularly yielding better subjective evaluation results than the ground truth on LibriSpeech test-clean and VCTK." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The authors have made an effort to present this work as a promising study; however, upon closer examination, there are numerous concerns that require substantial improvement.\n\n- **On Subjective Evaluation Results**: \n The results discussed as a strength in Figure 1 are fundamentally flawed due to the dataset disparity with other studies (e.g., NaturalSpeech3 [1] uses the Librilight dataset). This undermines fairness in comparison. To ensure meaningful comparison, the authors should replicate NaturalSpeech3, which currently appears to be a good model, and train on the same dataset.\n\n- **On Grouped Code Modeling**: \n While this approach has some merit as a method to reduce the sequence length given the codec model’s high 75Hz frequency, it is rather naïve and cannot be considered innovative. In fact, similar efforts have already been undertaken in existing research, such as [2], which the authors should have cited at minimum. Additionally, the method does not lead to significant improvements in either objective or subjective evaluations, suggesting that further refinements are needed.\n\n- **On Repetition Aware Sampling**: \n Although this method appears to address the traditional issue of repetition in models like VALL-E effectively, it is not particularly innovative. In language modeling (LLM) contexts, penalties for repetition have long been in use [3], making the lack of reference to these approaches surprising. While the authors’ method differs slightly from these established approaches, it would be necessary to compare with them to clarify the method’s effectiveness. Moreover, the existing application of repetition penalties in TTS contexts, as seen in [4], further accentuates this concern.\n\n- **On Ablation Studies**: \n There is a significant lack of ablation studies. The paper includes excessive unnecessary information; for example, the equations related to the model are redundant, and condensing this information would allow the inclusion of ablation studies directly in the main text. The limited experiments in the appendix also lack relevance. Ablations such as the presence or absence of prompts and dataset size variations are not particularly noteworthy, and their results are self-evident. More critical studies, such as comparisons with traditional repetition penalties or ablations involving Vocos (a major change from VALL-E), would have been more appropriate.\n\n- **On Baseline Comparisons**: \n Changing the decoder from VALL-E’s original to Vocos represents a major shift and warrants stronger emphasis in comparative experiments. Additionally, the fact that subjective evaluation is best when the group size is 1 makes it very challenging to establish differentiation from the baseline.\n\n- **Contribution to the Field**: \n The lack of code and weight release significantly diminishes the contribution of this study to the field.\n\n\n[1]: Ju, Zeqian, et al. \"Naturalspeech 3: Zero-shot speech synthesis with factorized codec and diffusion models.\" ICML 2024.\\\n[2]: Tang, Changli, et al. \"Salmonn: Towards generic hearing abilities for large language models.\" ICLR 2024.\\\n[3]: Keskar, Nitish Shirish, et al. \"Ctrl: A conditional transformer language model for controllable generation.\" arXiv preprint arXiv:1909.05858 (2019).\\\n[4]: Casanova, Edresson, et al. \"XTTS: a Massively Multilingual Zero-Shot Text-to-Speech Model.\" INTERSPEECH 2024." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "[Q1. Comparison with Low-bitrate Codec] Have you compared the grouped Code Modeling with low-bitrate Codec?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "They enhanced the baseline model, VALL-E, the first neural codec language model, by introducing repetition-aware sampling and grouped code modeling. While the baseline models are prone to issues like word repetition or omission, the proposed methods mitigate these problems and further improve model efficiency." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors proposed a neural codec language model for zero-shot text-to-speech, enhancing robustness by refining sampled tokens through repetition-aware sampling. They further improved robustness and efficiency by applying grouped code modeling, effectively reducing sequence length." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.\t[Grouped Code Modeling1] From an engineering perspective for neural codec language models, the proposed grouped code modeling could improve model performance and efficiency. However, grouped code modeling is already a well-known technique in language models, as seen in works like [MegaByte], [RQ-Transformer], and [Block Transformer].\n\n[MegaByte] Yu, Lili, et al. \"Megabyte: Predicting million-byte sequences with multiscale transformers.\" Advances in Neural Information Processing Systems 36 (2023): 78808-78823.\n\n[RQ-Transformer] Lee, Doyup, et al. \"Autoregressive image generation using residual quantization.\" Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2022.\n\n[Block Transformer] Ho, Namgyu, et al. \"Block Transformer: Global-to-Local Language Modeling for Fast Inference.\" arXiv preprint arXiv:2406.02657 (2024).\n\n2.\t[Grouped Code Modeling2] Additionally, [UniAudio] and [GPST] have already adopted a similar structure to sample RVQ tokens more efficiently. While there may be slight differences in implementation, they have the same goal\n\n[UniAudio] Yang, Dongchao, et al. \"UniAudio: Towards Universal Audio Generation with Large Language Models.\" Forty-first International Conference on Machine Learning.\n\n[GPST] Zhu, Yongxin, et al. \"Generative Pre-trained Speech Language Model with Efficient Hierarchical Transformer.\" ACL, 2024.\n\n3.\t[Grouped Code Modeling3] Notably, the classic sequence-to-sequence text-to-speech Tacotron also utilized the grouped spectrogram sampling through a reduction factor.\n\n[Tacotron] Wang, Yuxuan, et al. \"Tacotron: Towards end-to-end speech synthesis.\" arXiv preprint arXiv:1703.10135 (2017).\n\n4.\t[Grouped Code Modeling4] The recently proposed model [MELLE] also claims to predict multiple frames per step, accelerating inference and mitigating robustness issues associated with long-sequence modeling while maintaining strong performance. Moreover, MELLE has been shown to outperform VALL-E2. This hurts the contribution of the proposed method.\n\n5.\t[Repetition Aware Sampling] Recently, Flow-matching and MaskGIT-based text-to-speech models have adopted iterative sampling methods similar to repetition-aware sampling. It would be beneficial for the authors to discuss and compare repetition-aware sampling with these iterative methods, particularly against models like VoiceBox and E2-TTS.Specifically, I hope to see the comparison with VoiceBox and E2-TTS. \n\n6.\t[Weak Baseline] The authors only compared the model with VALL-E. However, VALL-E underperforms compared to VoiceBox, E2-TTS, DiTTo-TTS, and CosyVoice. \n\nI sincerely acknowledge the novel contribution of VALL-E in opening the door for neural codec language models; however, the novelty of VALL-E2 does not meet the standards expected for ICLR." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 4 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "- Why do this paper choose Byte-Pair Encoding (BPE) for text tokenization instead of using phonemes? How many BPE tokens are used in the model? Given that large datasets like LibriHeavy typically require thousands of BPE classes, while phoneme-based tokenization usually involves only a few dozen classes, how do you anticipate this choice impacts the model’s performance?\n\n- Including punctuation marks in modeling units could benefit text-to-speech (TTS) systems. I’m interested to know if the BPE units in this work incorporate punctuation marks. How might this decision impact the model’s performance?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "- VALL-E 2 achieving human parity in zero-shot TTS is a promising advancement, marking a new benchmark for text-to-speech systems. Its potential applications are particularly promising in assistive technologies for individuals with speech impairments.\n\n- The introduction of repetition-aware sampling and grouped code modeling is a simple but effective approach that enhances the model's stability and efficiency in generating speech. These methods could be easily adapted to speech-to-speech language models.\n\n- The paper demonstrates strong experimental validation with comprehensive evaluations on datasets including LibriSpeech and VCTK, showing clear improvements in robustness, naturalness, and speaker similarity​.\n\n- The technical explanations and results are clearly presented, making the contributions and performance enhancements easy to understand." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "VALL-E represents a breakthrough in neural codec language modeling for zero-shot text-to-speech synthesis. It can synthesize personalized speech from just a 3-second recording, while preserving the speaker's voice, emotion, and acoustic environment. VALL-E uses an autoregressive transformer to model coarse codec codes (1st group of EnCodec) and a non-autoregressive transformer to generate fine codec codes (2nd-8th groups of EnCodec). However, VALL-E faces two key limitations: 1) Stability: Random sampling during inference can cause instability, while small top-p nucleus sampling risks infinite loops. 2) Efficiency: Its autoregressive architecture is constrained by a fixed high frame rate, slowing inference.\n\nThe paper introduces VALL-E 2, which addresses the aforementioned issues with two innovations: Repetition Aware Sampling, which stabilizes decoding without increasing computational costs, and Grouped Code Modeling, which reduces sequence length and speeds up inference. These improvements make VALL-E 2 more robust, natural, and efficient in zero-shot TTS, achieving human parity for the first time on benchmarks including LibriSpeech and VCTK. VALL-E 2 can stably generate high-quality speech for complex sentences that are hard to read or contain many repeated phrases." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "While VALL-E 2 delivers remarkable improvements in stability and efficiency, it doesn't introduce the same level of paradigm-shifting innovation as the original VALL-E, which opened a new avenue for zero-shot TTS. VALL-E 2 focuses on refining and optimizing the existing framework, building on established concepts." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "NA" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "see above" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work is an extension of VALL-E, which aims to solve its two problems. 1. inference repetitions --> degrade performance 2. too long codec sequence for modeling --> degrade speed.\n\nSpecifically, they propose 1. Repetition Aware Sampling to remove repititions by accounting for token repetition in the decoding history, thus improve the synthesis quality, 2. Grouped Code Modeling to re-organize codec sequence into groups to shorten the length, thus improve the modeling efficiency.\n\nFrom my side, it is a good extension of VALL-E series and solve its practical issues. But from research perpective, this work does not convey much novelty or insights, especially given the high requirement of ICLR conference." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "see above" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024valle,\ntitle={{VALL}-E 2: Neural Codec Language Models are Human Parity Zero-Shot Text to Speech Synthesizers},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=0bcRCD7YUx},\nnote={under review}\n}" }, "abstract": { "value": "This paper introduces VALL-E 2, the latest advancement in neural codec language models that marks a milestone in zero-shot text-to-speech synthesis (TTS), achieving human parity for the first time. Based on its predecessor, VALL-E, this work introduces two significant enhancements: Repetition Aware Sampling refines the original nucleus sampling process by accounting for token repetition in the decoding history. It not only stabilizes the decoding but also circumvents the infinite loop issue. Grouped Code Modeling organizes codec codes into groups to effectively shorten the sequence length, which not only boosts inference speed but also addresses the challenges of long sequence modeling. Our experiments on the LibriSpeech and VCTK datasets show that VALL-E 2 surpasses previous systems in speech robustness, naturalness, and speaker similarity. It is the first of its kind to reach human parity on these benchmarks. Moreover, VALL-E 2 consistently synthesizes high-quality speech, even for sentences that are traditionally challenging due to their complexity or repetitive phrases. The advantages of this work could contribute to valuable endeavors, such as generating speech for individuals with aphasia or people with amyotrophic lateral sclerosis. See https://anonymous/valle2 for demos of VALL-E 2." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Zero-shot Text to Speech Synthesis", "Speech Generation", "Voice Cloning", "Language Modeling", "In-Context Learning" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/579fe66d05f6a3deb73edd5cbcfb8b8a4acf66e6.pdf" }, "presentation": null, "primary_area": { "value": "applications to computer vision, audio, language, and other modalities" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/5c1e23d71a23cebe5d23d6627c3332837403d792.zip" }, "title": { "value": "VALL-E 2: Neural Codec Language Models are Human Parity Zero-Shot Text to Speech Synthesizers" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "1. should we assume that $\\mu_k\\ge c$ for all $k$? The authors state that the optimal action is always $(m_1,\\dots, m_K)$ in line 211. It seems that this only holds when $\\mu_k\\ge c$ for all $k$.\n2. what is \\mu in Theorem 1. I did not find the definition.\n3. I am curious about how could the sample complexity in Theorem 2 gets rid of the dependence of N. Intuitively, even there is no noise (sigma = 0), for any algorithm, it still need at least $\\log N$ rounds to find the true $m_k$ by binary search. Is the dependence on $N$ hidden in $\\xi$?\n\n\nTypos:\nline 224: $a_{k,t}$ instead of $a_k$\nline 289: for large probability -> with high probability\nline 383: to played with -> to play with" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The work closes the sample complexity gap and narrow the regret gap for the MP-MAB problem. Although the techniques used in the proof are not particularly unique (mostly based on regular UCB and LCB), the conclusions are still very interesting and make sense.\n2. The work propose numerical simulation to show the advantages of their algorithms." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper revisits multi-play multi-armed bandit with shareable arm capacities problem. Improved on previous work Wang et al. (2022a), the paper proposes refined lower and upper bounds for both sample complexity and regret. For sample complexity, the authors propose a minmax lower bound, and give an algorithm that matches the bound. For regret, the authors provide both instance dependent and instance independent regret lower bounds, and find algorithms that match the bounds up to some model-dependent factors." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The writing is a bit poor. The paper contains many colloquial expressions, i.e., line 383 \"But if\", line 390, 403, 405 \"And furthermore\" \"And this\". \n2. The author states in the introduction that the algorithm has applications to LLM inference serving. I believe it’s necessary to provide some LLM-related experiments to support this statement." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1.\t$m_k$ is deterministic and well known beforehand as to how many pulls can be made in a round. However, there is a constant movement cost $c$ associated to an arm. In case of LLM query, the number of pulls is associated to the amount of query a server instance can handle.\n\nAre $m_k$ and moving cost $c$ dependent in this scenario? If so, how do this implication sit with all the theoretical proof, or do they have to be independent? A clarification on this would help the readers utilize the developed algorithms on many scenarios where the dependencies are crucial. \n\n2.\tWhy is ordering of plays of arm selection important in $a_t$, providing some details on it will avoid ambiguity around its objective of whether to maximize the resource utilization or to maximum capacity of the arm?\n\n3.\tAlso, with respect to movement cost, in the experiment setting, it has been assigned to an arbitrary value of 0.1 . Is there any fundamental reason for that? Also, how can they be evaluated in a practical scenario when they are also coupled with reward formulation? Adding some details around them can greatly improve the clarity of the work. \n\n4.\tIt would be nice to see how the experiments scale up with varying parameters like changing the $m_k$ and changing the movement cost etc ? This will help us understand the empirical performance of the algorithm much better.\n\nReference:\n [A] Xuchuang Wang, Hong Xie, and John C. S. Lui. Multiple-play stochastic bandits with shareable finite-capacity arms. International Conference on Machine Learning, ICML 2022." }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "•\tThe problem of Multi play multi-armed bandit problem is an interesting setting to study and improve the foundation of it as it pertains to main real-world settings including LLM inference serving. The work re-establishes that with emphasis on theoretical guarantees.\n\n•\tThe work provides theoretical improvements in sample complexity compared to the existing work on MP-MAB-SAC. It tends to close the sample complexity gap found in the previous work in Reference A\n\n•\tThe authors also provide a new Improved algorithm, PC-CapUL that performs much better than other existing algorithms and have a solid theoretical backing to it with proved theoretical Regret bound guarantees.\n\n•\tThe experiments cover the regimes where the number of arms is larger which predominantly requires more exploration to take place. The developed algorithm provides much better performance in terms of regret compared to other existing algorithms in this experimental setting.\n\nReference:\n [A] Xuchuang Wang, Hong Xie, and John C. S. Lui. Multiple-play stochastic bandits with shareable finite-capacity arms. International Conference on Machine Learning, ICML 2022." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper studies the multi play multi-armed bandit problem having shared arm capacity where the in each round, the learner gets to select the arm for a number of pulls capped by the capacity limit with the goal of maximizing the total reward at the end of the play. The authors propose a new reward function and develop a new algorithm PC-CapUL for this problem setting. The developed algorithm provides tighter bounds on sample complexity and regret in comparison to the existing works, efficiently balances exploration and exploitation. The work is applicable in resource allocation problem with capacity constraint scenarios such as LLM inference and many other real world scenarios." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "•\tThe experimentation design could have been done much better with the inclusion of better baseline comparison in addition to the algorithm found in Reference A . Also, utilizing a real-world dataset for evaluation would have further complemented these theoretical results.\n\n•\tThe readability of the paper could be much improved. Also, a brief intuitive explanation like a proof sketch could be added in the main text to help the reader get the intuitive logic and understanding of the proof techniques. \n\n•\tA more detailed theoretical comparative analysis like how regret fares against the regret of other algorithms would make the argument much stronger for the developed PC-CapUL algorithm. Moreover, having such a discussion would also help us uncover insights like how the regret bound behaves in different regimes.\n\nReference:\n [A] Xuchuang Wang, Hong Xie, and John C. S. Lui. Multiple-play stochastic bandits with shareable finite-capacity arms. International Conference on Machine Learning, ICML 2022." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "None." }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "See \"Weakness\" Section for questions.\n\n- I wonder whether the dependency on the cost parameter $c$ can be improved for the regret lower bound." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- This paper first considers this problem with scarce shareable arm capacities and proposes both lower and upper bound for both sample complexity and the regret bound.\n- Based on the parts that I checked, the proofs look correct to me.\n- The experiments are also conducted to show superior performance compared to the previous work." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper considers the problem of multi-play multi-armed bandits with scarce shareable arm capacities. Specifically, different from [Wang et al., 2022a], this paper considers the problem where $N\\geq \\sum_k m_k$ where $m_k$ is the capacity of action $k$. With a modification on the reward function, this paper proposes new sample complexity lower/upper bound that is tight as well as regret lower/upper bound for this problem. Specifically, the author claims that the sample complexity lower bound proven in this paper improves upon the one shown in [Wang et al., 2022a]. Empirical results are also shown to strengthen their theoretical findings." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- One main concern is the motivation of this paper to consider the case where $N\\geq \\sum m_k$. In this case, the problem seems to be easier (in the sense of algorithm design) since you will definitely explore each action sufficiently enough to figure out the exact $m_k$ while in the opposite case $N< \\sum m_k$, the problem seems to be harder since you need to decide the exploration amount for the suboptimal $k$. Can the authors explicitly justify the choice of studying the $N\\geq \\sum m_k$ case and why it is challenging compared to the previous case?\n- This also leads to the question about the comparison between the lower/upper bound shown in this paper and [Wang et al., 2022a]. While the authors claim better lower bound, I wonder whether the upper/lower bound are comparable in these two cases? Can the algorithm that is derived in this setting adapted to the other? Moreover, I am not sure why equation (5) is more reasonable since it makes sense to me to have the noise's variance larger when $m_k$ or $a_k$ is large.\n- As for the upper bound, the bounds in Theorem 5 seems to be suboptimal since it seems to be dependent on $\\frac{\\max_i \\mu_i}{\\min_i \\mu_i}$, which can be large.\n- I do not understand the lower bound argument shown in Theorem 4. When the cost $c=0$, then this ratio becomes 0, which is surely not informative. In addition, why is the ratio independent of $m_k$? Can the authors explain more on this?\n- Typos:\n - Line 223: it -> if\n - Line 224: a_k -> a_{t,k}?\n - Line 471: depended -> dependent \n - Line 751: missing lemma reference.\n - missing periods at the end of many theorem statements (e.g. Theorem 4,5,6..)" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Please try to solve the problems in weaknesses. In addition, since the improvement in results achieved in this paper mainly comes from the careful selection of UCB, I would like to know what kind of inspiration it will bring to future work? This is not necessary as the theoretical improvement itself is interesting." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The theoretical contributions are nontrivial. This paper shows tighter lower bounds, and then proposes new algorithms to match them. Furthermore, the experiments verified the theories." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper discusses the problem of the Multi-play Multi-armed Bandit Model with Shareable Arm Capacities (MP-MAB-SAC). It tightens the lower bounds for both sample complexity and the cumulative regret compared to the previous work. Besides, this paper proposes corresponding algorithms to match the lower bounds. Finally, the numerical experiments show that the proposed algorithms outperform the existing ones." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I have the following concerns: \n\n1. The writing quality of this paper falls below the standards required for publication in ICLR. Issues such as clarity, rigor, and basic grammatical correctness are prevalent. It appears that the authors did not thoroughly review the paper before submission. From a writing perspective, the paper remains in draft form: numerous typos, confusing notations, and grammatical errors hinder readability. For example, (1) in Lemma 2, $\\epsilon^{uE}$ should be $\\epsilon^{UE}$; (2) in the proof of Lemma 2 “Bourel et al., 2020” is even not cited; (3) in the proof of Theorem 1, which lemma is used here? Besides, this theorem should be proved more formally; (4) What is the first baseline “MP-MAB-SA” in the experiments?\n\n2. The explanations provided in the paper are insufficient. (1) In Section 1, more concrete examples of the model's practical applications are needed. (2) The claim that certain changes in settings make the model more suitable for LLMs requires stronger evidence. For instance, the movement cost $c$ (which is known to the learner) seems irrelevant. (3) The paper should provide a more in-depth analysis of the experimental results, going beyond mere statements of fact.\n\n3. The comparison with the previous work seems not fair. (1) Since $N \\ge M$ makes the learner only need to learn the capacity $m_k$, without needing to learn the rank of the arms, the learning task seems easier. (2) In lines 307~310, is there any evidence to show stability is getting better? Besides, I’m kind of confused about this result because the robustness v.s. regret usually has some trade-off, which means the increasing of stability may (not always) lead to the decreasing of performance." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024multiplay,\ntitle={Multi-play Multi-armed Bandit Model with Scarce Sharable Arm Capacities},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=0bcUyy2vdY},\nnote={under review}\n}" }, "abstract": { "value": "This paper revisits multi-play multi-armed bandit with shareable arm capacities problem (MP-MAB-SAC), for the purpose of \nrevealing fundamental insights on the statistical limits and data efficient learning. The MP-MAB-SAC is tailored for resource allocation problems arising from LLM inference serving, edge intelligence, etc. It consists of $K$ arms and each arm $k$ is associated with an unknown but deterministic capacity $m_k$ and per-unit capacity reward with mean $\\mu_k$ and $\\sigma$ sub-Gaussian noise. The aggregate reward mean of an arm scales linearly with the number of plays assigned to it until the number of plays hit the capacity limit $m_k$, and then the aggregate reward mean is fixed to $m_k \\mu_k$. At each round only the aggregate reward is revealed to the learner. \nOur contributions are three folds. 1) \\textit{Sample complexity:} we prove a minmax lower bound for the sample complexity of learning the arm capacity $\\Omega(\\frac{\\sigma^2}{\\mu^2_k} \\log \\delta^{-1})$, and propose an algorithm to exactly match this lower bound. \nThis result closes the sample complexity gap of Wang et al. (2022a), whose lower and upper bounds are $\\Omega(\\log \\delta^{-1})$ and $O (\\frac{m^2_k \\sigma^2}{\\mu^2_k} \\log \\delta^{-1})$ respectively. 2) \\textit{Regret lower bounds:} we prove an instance-independent regret lower bound $\\Omega( \\sigma \\sqrt{TK} )$ and instance-dependent regret lower bound $\\Omega(\\sum_{k=1}^K\\frac{c\\sigma^2}{\\mu_k^2} \\log T)$. This result provides the first instance-independent regret lower bound and strengths the instance-dependent regret lower bound of Wang et al. (2022a) $\\Omega(\\sum_{k=1}^K \\log T)$. 3) \\textit{Data efficient exploration:}we propose an algorithm named \\texttt{PC-CapUL}, in which we use prioritized coordination of arm capacities upper/lower confidence bound (UCB/LCB) to efficiently balance the exploration vs. exploitation trade-off. We prove both instance-dependent and instance-independent upper bounds for \\texttt{PC-CapUL}, which match the lower bounds up to some acceptable model-dependent factors. This result provides the first instance-independent upper bound, and has the same dependency on $m_k$ and $\\mu_k$ as Wang et al. (2022a) with respect to instance-dependent upper bound.But there is less information about arm capacity in our aggregate reward setting. Numerical experiments validate the data efficiency of \\texttt{PC-CapUL}." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Multi-play multi-armed bandit", "scarce sharable arm capacity", "regret bounds" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/1d82406c145f097170a10408c7d2e4d3de679e72.pdf" }, "presentation": null, "primary_area": { "value": "reinforcement learning" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Multi-play Multi-armed Bandit Model with Scarce Sharable Arm Capacities" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please address the concerns raised in the weaknesses." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1) The idea of using an LLM to separately generate queries, files, and trajectories, followed by a query-file verifier and trajectory verifier is neat.\n2) The paper addresses the problem of using the appropriate array of tools corresponding to the information relevant to a query well. \n3) The experiments are thorough." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a method for multi-modal agent tuning for tool usage and presents a dataset designed to train an agent for tool usage. The authors claim that their T3-Agent achieves significant improvements." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1) Verifying the output of an LLM by the LLM itself does not seem accurate. I am skeptical about the quality of the generated MM-Traj dataset. \n2) You need quantitative verification for the dataset. (It is not clear whether a user study involving a few people on 100 data points out of 15K would provide sufficient confidence.)\n3) Experimental results on the GTA benchmark are more promising than those on the GAIA benchmark. However, the overall performance of T3-agent is not superior to that of the other agents in comparison. Specifically, on the GAIA benchmark, the HF agent with GPT-4o performs twice as well as the T3-Agent.\n4) If the T3-Agent’s performance were clearly superior, I would be more optimistic about the generated dataset. However, the results seem to support my doubts about the dataset.\n\nMinor comment: In Tables 2 and 3, the best results should be highlighted in bold." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "Q: Could you please elaborate on how not using the final answer A aligns with your goal? Specifically, how does this choice benefit your approach to enhancing tool-usage capability?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. Innovation in Multi-Modal Interaction: The approach shows potential in pushing the boundaries of how agents interact with cross-modal data sources. By focusing on practical applications of tool usage, this work could offer useful insights into building agents that understand and respond to complex queries across various media.\n2. Comprehensive Methodology: The paper describes a well-structured experimental setup and provides a clear description of the multi-modal tuning process. This includes thoughtful considerations on data processing, model architecture, and task-specific tuning steps, making it easy to follow.\n3. Evaluation Metrics: The authors employ a diverse set of metrics to evaluate agent performance. This choice not only validates the model’s accuracy in the tasks but also emphasizes the practical utility of the proposed framework in real-world applications." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a novel approach for multi-modal agent tuning, aimed at enhancing agent performance in complex environments through better utilization of multiple data modalities. The authors propose a tuning framework designed to leverage cross-modal data (e.g., visual and text info) to improve agent task performance, with specific emphasis on tool usage within the agent's capabilities. Their T3 agent is multi-modal agent that can efficiently use tools to solve practical tasks by tuning a VLM as the controller. Evaluations over the various dataset show the significant improvement using their agent with closed as well as open source model. Additionally, they curated dataset using multimodal info having trajectory of various length for broader study. In this work focus is on the correct tool selection and code is given more importance than the widely used JSON schema. In summary, they generate data, tune the VLM and create dataset followed by leverage of tool agent to make use of tool in a better way." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Interpretability: While the approach demonstrates performance gains, the interpretability of results remains limited. Additional analyses, such as ablation studies or attention maps, would be beneficial to understand how each modality contributes to the decision-making process. For example paper first generate queries first without files before relevant queries, what is the impact if we don't do that, or this is based on some past work/observations?\n\n2. Scalability: The paper does not thoroughly address the scalability of the proposed method, particularly as the number of modalities or the dataset size increases. It would be beneficial to test how the method's performance and computational requirements scale with additional modalities or larger datasets. For example, experiments that measure latency, memory usage, and accuracy as more data is introduced could illustrate the framework's robustness and its viability in resource-constrained or high-throughput environments.\n\n3. User Study: To evaluate the practical usability of the framework, a small user study or qualitative feedback from users would provide valuable insights into the query handling experience. Specifically, gathering feedback on aspects like ease of use, perceived accuracy, responsiveness to complex queries, and the intuitiveness of the tool-usage process could highlight areas for refinement in real-world settings.\n\nMinor hints:\n1. sec5.6 typo.. \"wen based\"--> web based.\n2. sec 3.4- Author(s) mention details can be found in but missed cross-referencing it.\n3. cross reference missing at end of sec 3.4" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "I lean towards acceptance because the paper introduces a promising approach to enhancing the reasoning capabilities of multi-modal agents. The proposed data synthesis pipeline and the resulting MM-Traj dataset are significant contributions to the field, potentially advancing the state of multi-modal learning. However, the paper lacks a discussion on the T3-Agent's robust programming capabilities and does not explore how these might be improved. I would like the authors to comment on this aspect." }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The data synthesis pipeline introduces a scalable, automated approach to generating diverse, complex multi-modal data for tool usage scenarios.\n- Verification mechanisms embedded within the pipeline enhance data quality, resulting in a robust, comprehensive dataset.\n- With training on the MM-Traj dataset, the T3-Agent demonstrates significant performance gains, surpassing agents built on closed-source models like GPT-4 in certain benchmarks.\n- Ablation studies underscore the critical role of data verification in achieving top performance.\n- The paper includes detailed visualizations of the T3-Agent’s reasoning process." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a new approach for improving tool usage in multi-modal agents by fine-tuning a vision-language model (VLM) controller with synthesized tool-usage data. To overcome the limitations of traditional LLM-driven agents—such as reliance on prompt engineering and limited reasoning for tool usage across modalities—the authors create a three-step data synthesis pipeline. First, *Query Generation* uses GPT-4o mini to generate diverse, tool-aware prompts. Next, *File Generation* retrieves images from similar datasets and creates other files programmatically. Finally, *Trajectory Generation* employs a zero-shot agent using GPT-4o mini to solve tasks, capturing thought processes, code, and observations for each step. Quality is controlled through query-file and trajectory verifiers, also based on GPT-4o mini, producing a dataset called MM-Traj. The resulting agent, T3-Agent, uses the ReAct framework and the MiniCPM-V model trained on MM-Traj, enabling versatile tool usage across categories like web search, visual perception, image editing, and file handling. Benchmarks on GTA and GAIA demonstrate the T3-Agent’s significant improvements in tool usage over both untrained VLMs and other state-of-the-art LLM-driven agents." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The T3-Agent exhibits a gap in programming capabilities, which leads to lower accuracy in predicted answers compared to GPT-4o. \n- While the paper acknowledges the T3-Agent’s limited programming capabilities, it does not suggest potential improvements or outline future directions to strengthen this aspect.\n- The reliance on GPT-4o mini throughout the pipeline raises questions about biases and limitations from this closed-source model. Exploring alternative methods or open-source models could enhance transparency and address these limitations.\n- What is $p_i$ in Equation 2?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. In Figure 3, the sum of trajectories—214 + 14,273 + 8,740 + 2,520 + 1,242 + 697 + 199 + 202—totals 28,087, which exceeds the stated 20k tasks in the abstract. In addition, the paper mentions that only 15k files remain after passing the query-file and trajectory verifiers. What is the final size of the generated dataset?.\n2. Why does GPT-4o mini outperform GPT-4o in Table 2, specifically in the row with HF Agents on AnsACC and CodeExec, given that GPT-4o is expected to be more powerful than GPT-4o mini?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The proposed data generation pipeline first generates queries independently of any specific files, followed by producing relevant files to align with these queries. This approach allows the pipeline to create more diverse and expressive tasks, unrestricted by file format or quantity limitations.\n2. This work introduces the novel MM-Traj dataset, containing 20k diverse tool-usage data points, supported by a human study to demonstrate dataset quality. \n3. This work also performs an in-depth statistical analysis of the dataset and shows that MM-Traj offers broad coverage across various data modalities and knowledge requirements, as shown in Figure 3." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work introduces a multi-modal tool-usage data generation pipeline designed to finetune vision-language models (VLMs) for tasks requiring tool-usage reasoning. The pipeline consists of three primary steps. First, a large language model (LLM) is prompted to generate query tasks. Next, relevant images or files are retrieved or generated based on the specified query tasks. Finally, ReAct agents are employed to generate trajectories that address the query task problem, followed by an additional LLM prompt to verify the generated data.\n\nThis study also introduces the MM-Traj dataset generated through the proposed scheme and uses it to finetune the MiniCPM-V model to create the T3-agent. The T3-agent's effectiveness is subsequently assessed on the GTA and GAIA benchmarks, showcasing a 20% improvement in performance over untrained VLMs and achieving results comparable to other baseline agents, such as the Lego Agent, Warm-up Act Agent, and HF Agent." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. In Section 3.3, this work states, “for other needed files, we prompt GPT-4o mini to extend the file content and generate Python code to produce the files.” However, the methodology for file generation remains unclear. For example, if a file is required to be an MP4 video of Taylor Swift’s latest album, it’s uncertain how this content could be generated through Python code alone. Furthermore, if GPT-4o mini generates the Python code to produce such files, it raises concerns about data quality and how the model ensures that the generated content is not hallucinated.\n2. While including a human study to assess dataset quality is commendable, having only five experienced raters for a subset of the data may be too limited, potentially introducing biases based on individual preferences. Gathering feedback from a **larger pool of participants**, even with fewer data points per person, could strengthen claims about the dataset's effectiveness. Additionally, comparing MM-Traj to filtered-out data may not yield meaningful insight. Instead, comparisons with other established tool-usage datasets would likely provide more meaningful insights.\n3. The evaluation results in Table 3 reveal mixed outcomes, with the T3-agent performing significantly worse than other methods, such as HF Agent and Sibyl Agent, on the GAIA benchmark. What could lead to this performance discrepancy on the GAIA benchmark?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024multimodal,\ntitle={Multi-modal Agent Tuning: Building a {VLM}-Driven Agent for Efficient Tool Usage},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=0bmGL4q7vJ},\nnote={under review}\n}" }, "abstract": { "value": "The advancement of large language models (LLMs) prompts the development of multi-modal agents, providing a feasible way to solve practical tasks by using tools. In this paper, we propose a multi-modal agent tuning method that automatically generates multi-modal tool-usage data and tunes a vision-language model (VLM) as the controller for powerful tool-usage reasoning. To preserve the data quality, we prompt the GPT-4o model to separately generate queries, files, and trajectories, followed by a query-file verifier and trajectory verifier. Based on the data synthesis pipeline, we collect the MM-traj dataset with 20k tasks using 10 tools. Then, we build the T3-agent that uses MiniCPM-V as the controller Trajectory Tuning for Tool usage using MM-Traj. Evaluations on the GTA and GAIA benchmarks show that the T3-agent has achieved remarkable improvements and outperforms GPT-4 driven agents by 10%, showing the effectiveness of the proposed data synthesis pipeline that leads to better reasoning capabilities in tool usage." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Multimodal Agents", "Vision-language Model", "Tool usage" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/bd65e7681d49d1ec78af6ccfa3a9cf75786a47ac.pdf" }, "presentation": null, "primary_area": { "value": "applications to computer vision, audio, language, and other modalities" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Multi-modal Agent Tuning: Building a VLM-Driven Agent for Efficient Tool Usage" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": null, "comment": { "value": "Thank you for your response. To clarify, while I was able to reproduce the results in the current paper using the provided scripts, the result was different from the original DeepSTARR paper. Otherwise, I encountered no additional issues." }, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": null, "primary_area": null, "questions": null, "rating": null, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": null, "summary": null, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": null, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": null, "comment": { "value": "Dear friend,\n\nThank you for your feedback. We’re sorry to hear about the difficulties reproducing DeepSTARR’s results. Could you provide more details regarding the specific issues or errors encountered during your experiments? This will help us identify potential discrepancies and support you more effectively. For reference, our experiments were conducted primarily on NVIDIA A40 GPUs, so any information on your setup and configurations might also be helpful in diagnosing the issue." }, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": null, "primary_area": null, "questions": null, "rating": null, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": null, "summary": null, "supplementary_material": null, "title": { "value": "Thanks for your feedback" }, "venue": null, "venueid": null, "weaknesses": null, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": null, "comment": { "value": "Fail to reproduce the results of Deepstarr using the provided scripts." }, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": null, "primary_area": null, "questions": null, "rating": null, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": null, "summary": null, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": null, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. What’s the rationale behind the collection of tasks used in this study? they seem to be very similar in terms of task type and it will be more interesting to see more variation in tasks such as zero-shot mutational effect prediction and generative sequence modeling.\n2. The paper is presented as a benchmark suite but the introduction of GenHybrid seems to be the main emphasis throughout the results section. However, the details of such model is missing from the main text of the paper. What’s the main focus of this paper? \n3. Why not include ab initio models trained on tasks specific data and naive benchmark in this suite? The numers presented in the paper is without context can hardly can be used as a standardized benchmark for future methods." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. Through and comprehensive dataset curation for gLM evaluation that covers both short and long range tasks.\n2. A wide range on gFMs are benchmarked across various model architectures and parameter sizes. \n3. The introduction of new hybrid method that leverages both attention based models and state-space models and outperformed existing models in most of the datasets evaluated. \n4. In-depth analysis of the benchmarking results and provided insights into the current state of gFMs and their performance differential in various tasks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduced a benchmark suite for genomic foundation models(gFMs) called GeneBench that systematically evaluates gFMs on an wide array of datasets across a range of tasks for both short and long range sequence prediction. This work also presented a new method called GenHybrid that leverages both SSM and attention based model architectures." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Lacking on tasks beyond classification and regression. One of the most promising application of gFMs is to predict zero-shot mutation effects and generative modeling of genomic sequences. This benchmark effort misses in both aspects. Many mutation effect databases are available and a comprehensive curation of a benchmarking dataset will be of vast interest to the community. \n2. Lacking vertical comparison of different model architectures on model sizes and pre-training schemes. I understand this will be computationally costly but as a benchmarking effort, this is necessary to paint a more complete picture of the model landscape. \n3. Missing naive benchmarks and ab initio models for comparison. It has been shown in many recent studies that gFMs do not outperform ab initio models trained on task specific datasets. Adding both ab initio models and native benchmarks will be very important as a benchmark suite." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "See weakness" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- Benchmarked large number of tasks. \n\n- Compared all major genomic foundation models developed recently." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a benchmark framework for evaluating genomic foundation models. The authors gathered a large number of tasks from multiple existing papers for benchmarking. The tasks are classified into either long-range tasks or short-range tasks. A study that compares several existing genomic foundation models using the gathered tasks was performed. In addition, the authors proposed a hybrid approach that is supposed to work well for both short-range tasks and long-range tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The training in Pre-training is not clear. Should the optimization in Eq. (1) be actually involved in pre-training? The two categories of targets described in fine-tuning do not appear in Eq. (1) at all. \n\n- The manuscript appears to be prepared in a hurry, needing major cleaning up. For example, there is no mentioning what the abbreviations used in Figure 4 stand for and in the caption of the same figure, it is mentioned (a), (b), and (c), but based on the content I believe only (c) is there. As another example, in line 406, the authors mentioned they studied NT with different choices of number of parameters. However, there is no description of the results and respective discussion to offer any insights. As the last example, in Table 7, Caduceus is the second performing model, but the author said it was HyenaDNA in the text. \n\n- There is no description on how the proposed Genhybrid was trained. \n\n- The value of their main findings may be limited. Since attention-based models were trained using short-length sequences while convolution-based models trained using long sequences (to consider long context), it is expected to see the former is better in short-range tasks and the later has potential advantage in long-range tasks. \n\n- The effort in data curation is minimal. It looks to me simply pulling from previous works. Could the authors clarify if there is any additional processing or validation of the data. \n\n- Due to the limitation of the work pointed by the authors themselves, I do not see they can answer the last two questions summarized in the second paragraph in their introduction. Could the authors explain?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- Could the authors discuss how GeneBench differs from or improves upon existing genomics benchmarking?\n- Could the authors provide detailed input-output descriptions for some tasks (e.g. Genomic Structure Prediction)? \n- For visualization in Figure 8, it would be helpful if the authors add x-axis and y-axis labels to the heatmaps. Similarly, for Figure 10 and 11, what's the range of y-axis?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The paper offers a wide-range and detailed evaluation of GFMs and also provides concrete guidance for users on how to select models based on different tasks.\n- The paper provides clear classification of GFMs and benchmarking tasks.\n- Beyond benchmarking, the paper proposes a new model GeneBench based on the insights from the experiments, and achieves the best performance on most tasks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces GeneBench, a benchmarking suite specifically designed for evaluating Genomic Foundation Models across a wide range of genomic tasks. GeneBench includes evaluations of eleven GFMs on forty-four datasets, with tasks spanning various genomic regions and functions. This systematic benchmarking reveals insights based on the performance of GFMs across short- and long-range tasks. Furthermore, the paper proposes a new model that incorporates advantages from two types of models and demonstrates effective performance across all tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The paper belongs to a benchmarking paper, but it does not include comparisons with other existing genomic benchmarks (e.g., the length of input sequences, types of benchmarked methods, etc.). This limits the motivation why the research area needs this new benchmarking.\n- While the benchmark focuses on GFMs, it's better to have simpler baselines without pretraining, (e.g. CNNs). Including such models would provide a deeper understanding of the advantages or limitations of GFMs relative to classical models.\n- The paper doesn't provide sufficient descriptions on several tasks (e.g. Genomic Structure Prediction)" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. The authors are suggested to state the advantages of their method. For example, why a bioinformatian should use their proposed method instead of others?\n2. The authors should provide some biological insights.\n3. Some case studies can be provided. For example, how the proposed method can be used to facilitate biological findings." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The manuscript is well-organised and the experiments are relatively comprehensive." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This study introduces a comprehensive benchmark suite, GeneBench, for evaluating the efficacy of Genomics Foundation Models. The authors systematically evaluated several DNA tasks including coding region, non-coding region, and genome structure. They also provided some insights into the model design and model training." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Lack of biological insights" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We introduce GenBench, a benchmark suite for evaluating Genomic Foundation Models (GFMs). Based on our experimental insights, we propose GenHybrid, an effective SSM-attention hybrid model suitable for all tasks." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024genebench,\ntitle={GeneBench: Systematic Evaluation of Genomic Foundation Models and Beyond},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=0bswm093Yl},\nnote={under review}\n}" }, "abstract": { "value": "The Genomic Foundation Model (GFM) paradigm is expected to facilitate the extraction of generalizable representations from massive genomic data, thereby enabling their application across a spectrum of downstream applications. Despite advancements, a lack of evaluation framework makes it difficult to ensure equitable assessment due to experimental settings, model intricacy, benchmark datasets, and reproducibility challenges. In the absence of standardization, comparative analyses risk becoming biased and unreliable. To surmount this impasse, we introduce GeneBench, a comprehensive benchmarking suite specifically tailored for evaluating the efficacy of Genomic Foundation Models. GeneBench offers a modular and expandable framework that encapsulates a variety of state-of-the-art methodologies. Through systematic evaluations of datasets spanning diverse biological domains with a particular emphasis on both short-range and long-range genomic tasks, firstly including the three most important DNA tasks covering Coding Region, Non-Coding Region, Genome Structure, etc. Our results on GenBench has led to an interesting discovery: regardless of the number of parameters, the noticeable variation in preference between attention-based and convolution-based models for short- and long-range tasks could offer valuable insights for the future development of GFM. As a result, we propose a straightforward modified model called Genhybrid, which is an effective and efficient convolution-attention hybrid model suitable for all tasks." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "genetic foundation model", "benchmark", "hybrid model" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/af90bfafa536d15615226019c9450f07a06e60fc.pdf" }, "presentation": null, "primary_area": { "value": "datasets and benchmarks" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "GeneBench: Systematic Evaluation of Genomic Foundation Models and Beyond" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "The paper develops a simple way to distill the multi-aspect knowledge of MLLM to perform image classification using a student model. The experiments show some improvements however I still believe that the contributions of this paper are quite limited. Additionally, the baseline models selected in this paper are quite outdated. In summary, the overall technical novelty of the direct injection of knowledge from large models seems incremental.\n\n1. As shown in Tab 1, MLLMs perform badly in zero-shot classification on fine-grained image test datasets, how do we ensure that MLLMs provide correct answers across multiple aspects? \n2. There are questions regarding the task details when extending to object detection: should the input to the MLLM be the object within the box or the entire image? The entire image may contain multiple objects, and the MLLM's response may not be accurate." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1.This paper is written in a clear and straightforward manner, making it easy to quickly grasp the method's approach.\n2.The paper conducted a lot of experiments, and the figures and tables are well-organized.\n3.The authors claimed they are the first to offer a novel perspective on distilling multi-aspect knowledge regarding abstract and complex concepts. I have seen the author's efforts in the design of knowledge transfer." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper starts from the perspective of how humans classify images, where humans typically consider multi-aspects such as context, shape, color, and other features. Motivated by this, the author proposes a multi-aspect knowledge distillation method that utilizes Multimodal Large Language Models (MLLMs) to improve image classification performance. By querying, extracting relevant logits, and expanding the model's output dimensions, the method achieves the knowledge learning of visual aspects and abstract knowledge. This method enhances the performance of baseline models across lots of experiments, demonstrating the potential of multi-aspect knowledge distillation in computer vision and other tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The proposed method shows some improvement on some classic CNN-based models but lacks experiments on ViT-based models. \n2. In the knowledge distillation task, the comparison is only done with KD, lacking comparisons with other knowledge distillation methods [1,2].\n3. The improvement in object detection tasks is very limited in Tab7, and there is no comparison done on currently well-performing object detection methods. Object detection is inherently a more fine-grained visual task than classification. Still, the experiments in this paper do not demonstrate the method's effectiveness of multi-aspect knowledge distillation in detection.\n4. The explanation for the poor zero-shot classification performance of MLLMs is missing in Tab 1. Incorrect knowledge could also be distilled to the student model.\n5. Missing the training curve of MaKD Loss with the number of iterations. The visualization of t-SNE embeddings and the model's multi-aspect responses to a single image are presented in Fig 4 and 5. There is no overall evaluation of the model's responses to multi-aspect on the test dataset.\n\n[1] Decoupled Knowledge Distillation\n[2] Logit Standardization in Knowledge Distillation" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "I hope the author can better explain the novelty of the paper and the principle of how the algorithm is really effective." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The paper is well-written, with a clear and logical flow from the introduction through to the conclusion. The authors present simple ideas in a straightforward manner, making the paper accessible to readers from diverse backgrounds. The experimental setup is meticulously organized, with each step of the process described in a way that facilitates reproducibility. The authors outline the methodologies, datasets, and evaluation metrics in clear subsections, allowing readers to follow the experimental design intuitively." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a new knowledge distillation method. It performs multi-aspect knowledge distillation with the LLM and the MLLM. LLM is utilized to generate multi-aspect questions by using the class and prompt. It further adopts the MLLM to extract the logit for multi-aspect questions and obtain the probabilities corresponding to yes token. The student is optimized by the original cross-entropy loss and the distilled binary cross-entropy loss. Extensive experiments are conducted to demonstrates the effectiveness of the proposed method. It also extends to object detection task to evaluate the great potential." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1、\tLimited Experimental Setting：The experimental setting is narrow, which restricts the generalizability of the findings. The scale of datasets is small and may not be sufficient to demonstrate the robustness of the proposed method across different scenarios. Expanding the experimental scope to include more varied or challenging datasets such as the full ImageNet would significantly strengthen the paper.\n\n2、\tLack of novelty: The proposed method directly adopts the MLLM’s output logit to perform distillation. The principle behind this design is not fully demonstrated. Why MLLM can help improve the performance of student and what features support this? \n\n3、\tSome details are missing, and some experimental comparisons are not fair. The parameter number of the MLLM is larger than the teacher model in the traditional KD. It is questionable whether the improvement is due to the large number of parameters or the inherent properties of the MLLM itself. What will happen to the performance of the student model if only adopting the large vision encoder in MLLM. Some comparison to the traditional methods is not fair. The basic KD adopted in experiment in classification is too old and many improved versions should be used.\n\n4、\tThere is not the comparison to SOTA KD method in object detection and the baseline should also adopt the powerful setting." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "N/A" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. I noticed that even using random logits leads to performance improvements (Table 3(b)). Could you clarify the underlying reason for this result?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The core idea is simple but looks effective.\n2. The paper writing is fluent and easy to follow.\n3. The paper conducts experiments on six different fine-grained datasets and two different coarse-grained datasets. The results show that the proposed method achieves stable performance improvement, especially on the fine-grained datasets.\n4. The ablation studies and related visualization are comprehensive and insightful." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a multi-aspect knowledge distillation framework that uses MLLMs to improve model performance in visual understanding and detection tasks. By expanding the model’s output dimensions, the method distills multi-aspect logits that encapsulate diverse visual and contextual features beyond standard class labels. Extensive experiments on various image classification datasets, complemented by thorough ablation studies, underscore the framework's effectiveness and robustness." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The evaluation datasets in the paper are relatively small, and the model parameters appear insufficient in 2024 . Using ResNet18/34 as the primary model limits the assessment of the framework’s scalability. It would be valuable to test the framework on a larger dataset, such as ImageNet, and with a more complex model like ResNet101, to assess its effectiveness in a more challenging setting.\n\n2. The paper lacks comparisons with other knowledge distillation (KD) baselines, which would provide a clearer benchmark for evaluating the proposed method’s relative performance.\n\n3. The framework could explore additional ways to leverage the knowledge in MLLMs. For instance, distilling logits from the last token output by the MLLM after processing the input image may capture different aspects of visual representation.\n\n4. While Section 5.5 discusses training time and computational cost, the analysis might be incomplete. The time required for MLLMs to annotate the training dataset should also be considered to provide a more comprehensive assessment of computational demands." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "No" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. The approach to utilizing knowledge distillation is somewhat unclear. When are the multi-aspect logits extracted from the MLLM, and how are they incorporated into the model's training or inference objective?\n\n2. Given that GPT-4o generates the multi-aspect questions and that the MLLM has not seen images from each category, especially considering these categories are often long-tailed and fine-grained. Do you have any validation or filtering steps in place for the generated questions and responses, or have you considered comparing the generated questions to human-curated ones? Which types of generated questions contribute most to performance improvements.\n\n3. When generating responses to the multi-aspect questions for each image, has the potential hallucination issue within the MLLM been considered? How accurately can the MLLM (InternVL) answer these generated questions, and to what extent does the hallucination issue in InternVL affect the accuracy of its responses?\n\n4. For the object detection task, have you attempted to use other datasets, such as the larger-scale LVIS?\n\nI may reconsider my score based on your response to these issues." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The paper combines traditional network models, such as ResNet, with MLLMs to enhance accuracy in classification and detection tasks. \n\n2. It uses multi-aspect questions to extract knowledge from MLLMs, leveraging this knowledge to support classification. \n\n3. The experiments are comprehensive." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a novel approach to solving computer vision tasks, such as image classification and object detection, by enhancing conventional models' classification capabilities through knowledge distillation from Multimodal Large Language Models (MLLMs). The method involves expanding the dimensionality of the model's original logits, which improves classification accuracy. The paper provides numerous ablation experiments and conducts a thorough analysis of the results." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The approach to utilizing knowledge distillation is a bit unclear—are you applying this strategy during training, or is it only used in inference? Additionally, there seems to be a lack of consideration for hallucination issues that may arise with GPT-4o during the generation of questions and responses." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024multiaspect,\ntitle={Multi-aspect Knowledge Distillation with Large Language Model},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=0cBttXaOUK},\nnote={under review}\n}" }, "abstract": { "value": "Recent advancements in deep learning have significantly improved performance on computer vision tasks. Previous image classification methods primarily modify model architectures or add features, and they optimize models using cross-entropy loss on class logits. Since they focus on classifying images with considering class labels, these methods may struggle to learn various aspects of classes (e.g., natural positions and shape changes). In contrast, humans classify images by naturally referring to multi-aspects such as context, shape, color, and other features. Inspired by this, rethinking the previous approach from a novel view, we propose a multi-aspect knowledge distillation method using Multimodal Large Language Models (MLLMs). Our approach involves: 1) querying Large Language Model with multi-aspect questions relevant to the knowledge we want to transfer to the model, 2) extracting corresponding logits from MLLM, and 3) expanding the model's output dimensions to distill these multi-aspect logits. We then apply cross-entropy loss to class logits and binary cross-entropy loss to multi-aspect logits. Through our method, the model can learn not only the knowledge about visual aspects but also the abstract and complex aspects that require a deeper understanding. We primarily apply our method to image classification, and to explore the potential for extending our model, we expand it to other tasks, such as object detection. In all experimental results, our method improves the performance of the baselines. Additionally, we analyze the effect of multi-aspect knowledge distillation. These results demonstrate that our method can transfer knowledge about various aspects to the model and the aspect knowledge can enhance model performance in computer vision tasks. This paper demonstrates the great potential of multi-aspect knowledge distillation, and we believe it offers a promising direction for future research in computer vision and beyond." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Multi-aspect Knowledge Distillation", "LLM", "MLLM" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/1f3ea1a0729b4006e2948a7a22aa1bbe8967403f.pdf" }, "presentation": null, "primary_area": { "value": "transfer learning, meta learning, and lifelong learning" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/86c44b4be3af16d50dbb0cae78b18c60d2df1ada.pdf" }, "title": { "value": "Multi-aspect Knowledge Distillation with Large Language Model" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "1. What is the use-case for this work if it will take minutes to hours to generate an output?\n2. How do these results change if you have a faster edge device compared to a CPU?" }, "rating": { "value": 1 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "The paper aims to solve a very important problem, how to run very large models with billions of parameters on CPUs or machines with no CUDA." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces a technique to run 70B LLM on CPU based (edge) devices. The system uses a tensor parallel framework to distribute attention heads across multiple nodes. The authors then perform an All-Reduce latency analysis, claiming that latency, not bandwidth, is the main bottleneck for all-reduce. The authors then describe a sliding memory scheduler, followed by experiments where they show the performance of their system." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Thank you for submitting your work to ICLR. I believe the current version of the paper has many shortcomings which I will describe here in detail:\n\n1. The paper needs thorough proof-reading. Some examples:\n- by adaptively partitioning model--->a model or models\n- with token latency increases-->increasing\n- Write in Active voice to avoid odd sentence structures such as :\n- \"Constrained by the high link latency, a star-based allreduce algorithm is implemented\"\n- \"We design a TPI-LLM\"--->We design TPI-LLM\n- \"power collaborate\"--->to collaborate\n- \"which dominate inference time\"---> \"what dominate\"\n\n\n2. I did not really understand the example on p.5 \"For example, the path from device h2 to h1 via h8\nfollows the route h2 → r2 → r9 → r8 → h8 → r8 → r9 → r1 → h1, resulting in a total link latency of 16τ , where τ is the per-hop link latency\"\n\n\n3. The Sliding Window Memory Scheduling is very similar to PipeSwitch (https://www.usenix.org/conference/osdi20/presentation/bai). The only main difference being that you are swapping in/out from Disk to/from device memory. This is in many ways also similar to memory page to disk swapping.\n\n4. Starting with the results for the swapping. Having an OOM is probably better than having 26.1s/token. For a 100 tokens output, you need to wait for roughly 30 minutes. This is an output of about 75 words as per OpenAI (https://help.openai.com/en/articles/4936856-what-are-tokens-and-how-to-count-them), i.e., one paragraph. Why would a user want to tolerate this? If the user chooses to use the fastest model (the Llama2-3B), they will wait for a bit less, about 3.3 minutes. I am not sure if there is a use-case for such a slow running LLM.\n\n5. Moving to the networking results in 4.2, I think the authors are drawing the wrong conclusions. The computations are just super slow in this case that the network is not really a bottleneck in the computations. I think a better experiment would be to try the system with proper edge GPU/TPUs, e.g., Google's Corel, NVidia's Orin, or NVidia's Nano. Right now, I believe that what we are seeing is the result of just a very slow computation. You can already see that in a real world scenario, things are much worse in the real case study where the latency in Table-3 is multiple times compared to Table-1." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "When the author discusses opportunities, I have several points of confusion:\n\n1. Why does the author believe that the communication proportion of tensor parallelism is high, but the overall inference time is reduced due to parallel computation? Is this conclusion drawn from the results in Figure 1-(b)? Has the author considered the synchronization issues among multiple edge devices in tensor parallelism?\n \n2. Why is the total time of tensor parallelism less than 100% in Figure 1-(b)? Does the author want to use the pipeline parallelism as a baseline to illustrate the superiority of tensor parallelism among edge devices?\n \n3. In Figure 1-(c), why is the memory footprint of each device in the TPI-LLM framework proposed in this paper the same? If tensor parallelism is used, it should decrease as the number of devices increases." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "This paper addresses the emerging challenge of deploying large language models (LLMs) on edge devices, which is a novel and increasingly relevant problem in the era of edge computing and privacy concerns. It proposes a new approach to tensor parallelism that is specifically tailored for low-resource environments, which is an original contribution to the field.\n\nThe authors combine concepts from distributed computing, memory management, and parallelism to create a system that is both memory and compute-efficient. The sliding window memory scheduler and the star-based allreduce algorithm are creative solutions that address specific pain points in edge device inference.\n\nThe work has significant implications for the future of edge computing, as it enables the deployment of LLMs on devices that were previously considered incapable due to resource constraints." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents a novel tensor parallel inference system named TPI-LLM, designed to efficiently serve large-scale language models (LLMs) on low-resource edge devices. The system addresses the challenges of limited computing power, memory, and bandwidth on edge devices by introducing a sliding window memory scheduler that dynamically manages layer weights during inference, overlapping disk I/O latency with computation and communication. This approach allows larger models to run smoothly on memory-limited devices while keeping sensitive raw data local to the user's devices, enhancing privacy.\n\nThe key contributions of the paper are as follows:\n\n1. Tensor Parallelism on Edge Devices: The paper argues for the effectiveness of tensor parallelism over pipeline parallelism on low-resource edge devices and presents TPI-LLM as a compute- and memory-efficient solution for serving 70B-scale models.\n \n2. Sliding Window Memory Scheduler: A memory scheduler is introduced to asynchronously load and unload layer weights, which enables the inference of larger models on devices with limited memory by overlapping disk I/O with computations and communications.\n \n3. Star-Based AllReduce Algorithm: The paper identifies link latency as the main issue in communication and implements a star-based allreduce algorithm to reduce latency, outperforming ring- and tree-based methods commonly used in high-latency networks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I think that although this paper is a bold attempt at the edge of LLM serving, most of the solutions provided are based on the application of past works. Firstly, whether it is tensor parallelism, sliding window, or the star-based algorithm, they are all proposed in existing and relatively mature works. The author's approach to using these methods to optimize edge LLM serving is similar to theirs, hence the paper's contribution lacks a certain level of innovation.\n\nFurthermore, I think there are some flaws in the author's logic when explaining the motivation and opportunities for the research. I only learned from the first sentence of the abstract that the significance of serving large models at the network edge lies in data privacy-preserving. In the first paragraph of the introduction, I learned that if edge devices must be used for LLM serving to ensure user privacy and security, then more edge devices will have to be used in a distributed collaborative serving due to the limitations of computing and storage resources. However, the title of the second paragraph is \"Observations and Motivations,\" but the content only contains observations (which should only contain them without motivations). Therefore, I suggest that the author optimize the logic of explaining the research motivation and opportunities. The scope of privacy security issues is too large, so that the research motivation seems slightly pale. Can it be combined with some more specific downstream tasks or application scenarios?\n\nFinally, the experimental results shown in Table 1 shows that the latency of serving large models on edge devices is much higher than that in the cloud, with TTFT reaching above the second level, and the throughput is far from comparable to that of the cloud. Does this indicate that the research motivation of the paper only considered privacy protection and neglected performance issues? Although the experimental results are already much better than Galaxy. This issue is also a huge challenge that all edge LLM serving inevitably faces." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "- This technique might also be useful even in the cloud server settings, especially when there is not enough GPUs, can the sliding window memory management can help avoid the OOMs. Any thoughts in that direction or future recommendations maybe?\n\nHowever, following are the concerns/questions.\n\n### Major concerns\n- The proposed framework has two major operational changes that are applied on the pre-trained LLM, that are 1) distribution of attention heads across nodes, ii) the memory management through the sliding window approach. Given these two changes, the paper does not discuss the performance implications with and without the proposed framework. It is important to guarantee that the performance remains same. \n - Given that, it is recommended to show that the performance (atleast in the best case settings on atleast one model) remains unchanged with or without the TPI-LLM.\n- Figure 5 is more concerning in the following perspectives.\n - No impact on bandwidth: There is no surprise on the lack of impact of increasing the bandwidth since the sliding window size is defaulted to 2. However, we can only learn about the impact by increasing the window size. There is no such ablation in the paper that shows a best combination of the number of devices, maximum possible sliding window size, bandwidth, available memory on each of the devices. \n - Recommendation is to provide a thorough feasibility study on the combination of the above variables to clearly understand the impact of the bandwidth. In fact on that note, there is no clear analysis on the maximum possible sliding window size for a given hardware configuration of the master/worker nodes. Therefore, the feasibility study can be preceded by maximum window size in order to limit the number of combinations to be studied.\n - Increasing the number of devices/cores reduces the token latency (from first two sub-figures in Figure5) is not a true statement and is really vague. Those plots are shown simply for 8 devices, if the number of devices are kept increasing, after a point we see diminishing returns of parallelism. That is where the communication dominates and hence to diminishing returns. Without having proper study `increasing the devices reduces the latency` are not valid claims. \n - The recommendation is to conduct a quick roofline analysis to substantiate the claims or remove the controversial parts. \n- There are a few limitations on what the proposed framework can offer. They are as follows.\n - There is a security and privacy concern, this framework should not run any device connected in the same Wi-Fi network unless there is a prior consent. It is not stated in addressed in the paper and hence please clearly state the limitation or the constraints under which the framework can operate.\n - The star configuration can not scale to large number of nodes/devices. It probably can be extended to scale in a hierarchical-star configuration etc, but that is not the scope of the paper and hence this needs to be state clearly as a limitation. There are real-time use-cases in the resource constrained edge scenarios where the number of devices is high, which leads to failures of a single master node in star config.\n - There is probably an unstated assumption in the paper that the data gets generated on the master or stays centrally on the master node/device. However, there is a high chance of the worker nodes/devices having user specific data. It appears that the proposed framework does not handle data that is generated on all the other devices. If that is not the case, please clarify how that data is handled on each of the workers? Otherwise, this limitation should be stated. \n\n### Minor concerns\n- In Figure 4, Time steps 7 and 8 have the same memory window, why is it the case, it is a bit confusing to understand. Is the memory peaked and hence the window does not slide or something else? Please provide clarifications or amend the figure to make it clear.\n- Tables 1 & 2 provide comparisons for TTFT, latency and memory usage, etc between with and without the use of memory scheduler. However, when the memory scheduler is disabled, it is not clear how those numbers are attained. How were they measured? by using accelerate, vanilla Klonet or Galaxy or llama.cpp or others?\n - Please add those details on how the stats were measured and the underlying frameworks. Ideally benchmark comparisons against the best possible frameworks/methods is a common practice?\n- Section 4.3 states the `comparison with benchmarks` Ideally those (a. Standalone, b. Model Parallelism (MP) and c. Galaxy) are SOTA methods/frameworks (in this case). Are they not? Why are they called benchmarks? Clearly there is benchmarking of TPI-LLM against those other things. \n - Recommendation is to please rephrase in order to convey the message so that the confusing the reader can be avoided." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "Following are the strengths:\n- The paper is well-written, easy to follow and understand the concepts presented.\n- The paper tries to addresses the critical problem of LLM inference on the edge devices.\n- The paper discusses the existing frameworks in this space and positions within the body of knowledge.\n- Achieving the faster TTFT latency using 3k LOC and for large models is always interesting and of use for broader audience.\n- Although the paper uses multiple devices connected in the same network (Wi-Fi/cable connected), it is understandable that there are use-cases where a house will have multiple edge devices all of which might operate in-tandem, there maybe business orgs that have edge devices that don't have permission issues to run LLMs on multiple of such devices. For such use-cases this is a significant contribution.\n- The paper proposed an easy to use technique of overlaying the data fetch during communication step in LLMs in the proposed sliding window strategy." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a new framework called TPI-LLM for model serving on low resource edge devices using tensor parallelism and memory scheduling. The proposed frameworks shows better performance over the SOTA frameworks in terms of latency to predict the first token and the overall token latency." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Please follow the questions section, there is a cohesive list of weaknesses and the corresponding questions that are to be addressed." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please see questions from weaknesses." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1 - The paper provides extensive theoretical analysis.\n\n2 - The proposed approach is evaluated in both emulations and real-world testbeds against state-of-the-art baselines.\n\n3 - The performance is significant, especially on edge devices with limited resources." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes TPI-LLM, a tensor parallelism-based distributed LLM inference system tailored to low-resource edge devices. Unlike inference tasks on servers, TPI-LLM keeps user data securely on user (edge) devices during inference. By analyzing the network latency bottleneck and memory limits, TPI-LLM leverages a star-based all-reduce algorithm to facilitate distributed inference and employs a sliding window-based memory scheduler to reduce inference memory footprints. Experiments on both emulations and real-world testbeds show that TPI-LLM outperforms existing baselines by providing lower TTFT, token latency, and peak memory footprints." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1 - In Section 2, Q1 is somewhat ambiguous. First, isn't tensor parallelism already a form of model parallelism? And second, even on low-resource edge devices, can we combine and use both parallelism techniques instead of simply abandoning one?\n\n2 - In Section 3.2, the example network topology here is star-based (Appendix A.7). Given this star-based topology, a star-based all-reduce scheme indeed would be most efficient. Is this a common network topology for all edge scenarios?\n\n3 - Figure 4 may need further detailed descriptions. The sliding window is not very clear in the figure. For example, in Time 7 and 8, why would you prefetch Blocks 6, 7, and 8 so early when it's still far away from actually using them? Isn't that prefetching too early causing memory waste?\n\n4 - Note that the layer-wise parameter offloading is not new. Many popular frameworks, such as DeepSpeed-Inference, Mixtral Offload, and Llama.cpp, support this offloading scheme. How does the proposed TPI-LLM differ from existing offloading techniques?\n\n5 - Evaluation lacks sensitivity analysis on the memory sliding window size. Why would you pick a window size of 2?" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "This work can serve 70B-scale LLMs efficiently using multiple edge devices with limited computing power, memory, and bandwidth." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024tpillm,\ntitle={{TPI}-{LLM}: Serving 70B-scale {LLM}s Efficiently on Low-resource Edge Devices},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=0cadcLKbt7},\nnote={under review}\n}" }, "abstract": { "value": "Large model inference is shifting from cloud to edge due to concerns about the privacy of user interaction data. However, edge devices often struggle with limited computing power, memory, and bandwidth, requiring collaboration across multiple devices to run and speed up LLM inference. Pipeline parallelism, the mainstream solution, is inefficient for single-user scenarios, while tensor parallelism struggles with frequent communications. In this paper, we argue that tensor parallelism can be more effective than pipeline on low-resource devices, and present a compute- and memory-efficient tensor parallel inference system, named TPI-LLM, to serve 70B-scale models. TPI-LLM keeps sensitive raw data local in the users' devices and introduces a sliding window memory scheduler to dynamically manage layer weights during inference, with disk I/O latency overlapped with the computation and communication. This allows larger models to run smoothly on memory-limited devices. We analyze the communication bottleneck and find that link latency, not bandwidth, emerges as the main issue, so a star-based allreduce algorithm is implemented. Through extensive experiments on both emulated and real testbeds, TPI-LLM demonstrated over 80\\% less time-to-first-token and token latency compared to Accelerate, and over 90\\% compared to Transformers and Galaxy, while cutting the peak memory footprint of Llama 2-70B by 90\\%, requiring only 3.1 GB of memory for 70B-scale models." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "DML Systems", "Edge LLM Serving", "Tensor Parallelism", "Memory Scheduling" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/06e1b42fa6edd155c508793b0d6c1310d7624497.pdf" }, "presentation": null, "primary_area": { "value": "infrastructure, software libraries, hardware, systems, etc." }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "TPI-LLM: Serving 70B-scale LLMs Efficiently on Low-resource Edge Devices" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "Is it possible to set the order of ellipsoids? Or how complicated would it be to extend this framework to allow user to set the order of ellipsoids?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "This paper is well motivated, formulated, written, and evaluated. \n\n1. The injection of ellipsoid information is achieved through cross attention. This allows the ellipsoids to be unordered set. i.e. user doesn’t have to specify the order or ellipsoids; model also decides the order of ellipsoids. \n2. The formulation of ellipsoid token is effective and is easy to be extended to different conditions other than secondary structure. (such as hydrophobicity)\n3. Thorough analysis of ellipsoid consistency, including both geometric and probabilistic metrics.\n4. Fig4: The comparison on designability/diversity with baseline methods are done as comparing Pareto frontiers with varying sampling temperatures/guidances. This clearly shows the tradeoff between the methods and the performance improvement from baselines. I found this analysis insightful and believe other papers arguing increased performance could benefit from a similar evaluation scheme.\n5. The practical use-case of the method is shown in Section4.3 flexible conditioning.\n6. The controllability is greatly improved from Chroma (Table1). Accuracy and coverage is very impressive." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a framework to generate protein structures by conditioning on layouts specified through 3D ellipsoids. The conditions include location, size, orientation and secondary structure. These conditions are injected to flow-base protein generative models via proposed cross-attention and update modules. It shows greatly improved controllability and designability over baselines." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "How is sequence design performance? What I understand is that the designability is solely based on the generated structure (i.e. generated sequence is discarded). Can you also present co-design designability value as in MultiFlow paper?\n\nOther than that, I did not find any major weaknesses from the paper. However, the ablation study can be improved. Most of the model ablations are based on guidance strength, but I am also curious about ablation study on (i) ellipsoid segementation cutoff (5A currently), (ii) allow residue token to update ellipsoid token or not. For (ii), explaining the reasoning behind the design choice may suffice." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. **Ellipsoid Representation**\n\n a. Choosing the Number of Ellipsoids (k):\n - *Training:* Is k determined by the structure of the training protein? If so, what is the distribution of k in natural proteins?\n - *Evaluation based on the statistical model:* The authors appear to have used a fixed k=5 in the experiments. How was this number chosen? Have the authors tested other k values?\n\n b. Number of Residues per Ellipsoid:\n - The current representation specifies the number of residues in each ellipsoid, but the authors show that this number directly depends on ellipsoid volume. Could specifying the number of residues be redundant, and might removing this constraint provide the model more flexibility in generation? Have the authors examined the impact of residue count on amino acid (AA) prediction?\n\n2. **Invariance Cross-Attention (ICA) Layer Design**\n\n a. The authors separately model the SE3 features (E_k) and scalar features (e_k) of ellipsoids using the proposed ICA and transformer to achieve SE3 invariance in the local frame. Has the team considered alternative approaches, such as modeling ellipsoids as “pseudo frames” with SE3 and scalar features and simply using IPA to update ellipsoid and residue features together?\n\n b. In Algorithm 1:\n - Could the authors clarify the *PosEmbed* used in line 223? Does it include distance, angles, or local coordinates?\n - Could they also explain why the query uses un-updated $s$, while the key and value use $a$, which incorporates current ellipsoid information?\n\n3. **Results in Table 2 (natural proteins):** The model, even with the strongest guidance, tends to overestimate helices in proteins. Additionally, the authors did not present designability results, which, based on Figure 16, may be compromised with strong guidance. Could the authors elaborate on the model's performance in addressing the “overrepresented helix problem,” the trade-offs with other metrics, and its overall comparison to models like *RFDiffusion*?\n\n4. In self-conditioning (line 290), they propose supplying interpolated conditions to both conditional and unconditional models, suggesting this improves “designability and ellipsoid adherence for all $\\lambda$ values.” However, no ablation studies were provided to verify this claim." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "**[Clarity & Quality]**\n- The manuscript is well-written, with a thorough introduction and background information on protein structure generation, spatial conditioning, and flow matching for data generation. Overall, it provides a smooth reading experience.\n- The paper is of good quality, presenting clear mathematical foundations grounded in current techniques for protein modeling, diffusion-based generation, and guided sampling.\n- The problem is well-defined, and the authors designed several experiments to evaluate model performance in 1) following conditioning, 2) improving general performance, and 3) demonstrating practical use in flexible conditioning, with both quantitative and qualitative comparisons.\n\n**[Significance]**\n- The spatial conditioning approach using ellipsoids is intuitive for practical applications and has potential implications for the utility of protein generation models.\n- The proposed methods appear generalizable to various spatial conditioning scenarios. (*However, this paper focuses solely on secondary-structure conditioning.*)\n\n**[Originality]**\n- The authors introduce a novel conditioning modality using spatial ellipsoids for protein generation, along with a new layer, \"invariant cross-attention,\" to integrate this information." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work extends Multiflow to accept spatial conditioning of secondary structures via 3D ellipsoids, aiming to improve control in protein generation and reduce the overrepresentation of alpha-helices in current generative models. Building on Multiflow’s architecture, the authors address two main challenges:\n\n1. Integrating and updating ellipsoid conditioning with structure embeddings with minimal modifications: they introduced an *Invariant Cross Attention* module to update residue embeddings while preserving local SE3 invariance.\n2. Implementing an effective conditioning approach for flow-matching models: they used classifier-free guidance to interpolate flow vectors across translation, rotation, and amino acid spaces, and employ self-conditioning to refine predicted structures\n \nExtensive experiments show that the proposed model can faithfully follow spatial conditioning, resulting in greater diversity, novelty, and improved secondary structure composition. This improves Multiflow by generating proteins with secondary structures more similar to natural proteins. \nOverall, this work presents a straightforward approach to control protein generation, enhancing Multiflow's diversity, novelty, and secondary structure accuracy." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "**[Clarity]**\n- Some design choices and model details lack clear explanations or in-depth examination (see Q1, 2, and 4).\n\n**[Soundness]**\n- Performance on Natural Proteins: While the authors demonstrate high designability at fixed helicity levels on synthetic data, it isn't as clear if these benefits hold for natural proteins (see Q3).\n\n**[Significance]**\n- Scope: The current methods are examined only on Multiflow and for secondary structure guidance. Their practical impact on other protein generation models and types of spatial conditioning (e.g., domains, hydrophobic cores) is not extensively explored." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 4 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "- Line 355: How is the length between ellipsoids determined/sampled? Also, consider an ellipsoid with beta strand annotation, how is the length between each stranded segment determined (particularly if a strand ellipsoid is formed by segments that are distant in sequence but close in structure)? \n- It would be helpful if the authors could provide ablation study results on the effect of self-conditioning, particularly the two self-conditioning schemes described in line 290.\n- Figure 9: What is the linear fit and statistical significance in both cases?\n- During training, is the ellipsoid conditioning information always provided, or only provided for a percentage of time?\n- Line 367: Why is a structured residue considered as “covered” if it is inside at least one ellipsoid instead of inside the ellipsoid it is assigned? \n- Table 1: Could the authors provide some intuition on why over-guidance (\\lambda > 1) performs better than the conditional model itself (\\lambda = 1)?\n- Table 2: Does “PDB proteins” correspond to the validation dataset or does it include the training dataset?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "The paper is well written with clean visualizations demonstrating methods and results. The concept of utilizing ellipsoids as conditions for protein generation is interesting and novel, providing a bridge between protein-level conditioning and atom-level conditioning. In addition, the authors propose an effective Invariant Cross Attention module for integrating ellipsoid conditioning and demonstrate success at achieving SOTA performance on the Pareto frontier of designability and diversity." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents ProtComposer, which is a fine-tuned model from MultiFlow to support conditional generation based on 3D ellipsoids, showing success at controllable design and achieving SOTA on the Pareto frontier of designability and diversity." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I have no major concerns about this paper. However, it would be helpful if the authors could elaborate on the applicability of the ellipsoid-based conditioning approach on practical protein design tasks. How would it help with or facilitate the protein design process?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "I am assuming that the user can redefine K during each trial as they see fit, though studies on guidelines in choosing K would be helpful. In situations where the user has only a vague idea of what they are looking for (an unfortunately common occurrence), having a guide on where to start would be beneficial." }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "It addresses an important and relevant area. The results are overall quite impressive as well. Additionally, showing the ability to work with handcrafted ellipsoid frames or the more easily scalable ML generated frames shows the practicality of ProtComposer.\n\nI really like the point about using simple ML models for ellipsoid sampling as compared to NNs. I would like to see some stronger analytical or experimental justification of the claim though." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Introduces ProtComposer, a generative model for proteins. ProtComposer seeks better control over the shape of generated proteins, as well as allow for greater novelty in the generation of proteins. Notably, the user has to choose between control or novelty, ProtComposer does not achieve both simultaneously. These tasks are accomplished by the introduction of 3D ellipsoid frames to guide the generation of proteins by the pre-existing Multiflow algorithm. \n\nWithout a pre-existing metric the authors feel sufficiently quantifies compositionality, they introduce their own metric. It is shown that ProtComposer outperforms existing methods in this metric." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Since a custom metric is introduced in this paper and then used as justification for the performance of the model, a section (either in the main paper or the appendix) justifying this metric would be nice. Showing comparisons to other pre-existing metrics, performance of many other algorithms under this metric, or stronger domain justification would strengthen the meaning of the results. I did not reject over this since the metric intuitively looks good, but further empirical or analytical support would be nice.\n\nThe formatting of the extra results in the appendices results in very odd page layouts, where some pages are blank, others have odd whitespace gaps, etc. Adjustment to make these pages more presentable would be nice." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We develop a framework to generate protein structures conditioned on spatial protein layouts that are specified via a set of 3D ellipsoids." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024protcomposer,\ntitle={ProtComposer: Compositional Protein Structure Generation with 3D Ellipsoids},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=0ctvBgKFgc},\nnote={under review}\n}" }, "abstract": { "value": "We develop ProtComposer to generate protein structures conditioned on spatial protein layouts that are specified via a set of 3D ellipsoids capturing substructure shapes and semantics. At inference time, we condition on ellipsoids that are hand-constructed, extracted from existing proteins, or from a statistical model, with each option unlocking new capabilities. Hand-specifying ellipsoids enables users to control the location, size, orientation, secondary structure, and approximate shape of protein substructures. Conditioning on ellipsoids of existing proteins enables redesigning their substructure's connectivity or editing substructure properties. By conditioning on novel and diverse ellipsoid layouts from a simple statistical model, we improve protein generation with expanded Pareto frontiers between designability, novelty, and diversity. Further, this enables sampling designable proteins with a helix-fraction that matches natural proteins, unlike existing generative models that commonly oversample conceptually simple helix bundles." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "protein design", "diffusion model", "controllable generation", "drug discovery", "proteins", "biology" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/150a497084667927c27cc7887351d629a94d394c.pdf" }, "presentation": null, "primary_area": { "value": "applications to physical sciences (physics, chemistry, biology, etc.)" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "ProtComposer: Compositional Protein Structure Generation with 3D Ellipsoids" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- For the procedure described in the Figure 1b caption on removing unimodal influence: why subtract out the unimodal contribution? Why not learn a regression directly from the unimodal contribution itself, i.e., the predictions of $r$?\n- Can it be said that the IB-audio and IB-video unimodal representations described on lines 223-224 are not truly unimodal, since they are extracted from a model that was trained with multimodal inputs? Then, they reflect correspondences between language and vision.\n- Figure 3 second row, middle two columns: Why does the green bar not have $\\wedge *$ for SV? It seems significantly higher than both light green bars. Why does the green bar have $\\wedge *$ for MT? It only seems significantly higher than on light green bar." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- This work is novel in that it is the first to use fMRI. But other works also take similar approaches to identifying multimodal processing (see weaknesses).\n- Findings of the difference between cross-modal and jointly-trained models with respect to regions is novel\n- These findings have neuroscientific significance\n- Appropriate random baselines are used to give context to alignment numbers is given" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "- Existing work uses unimodal models to identify language and vision processing pathways in the brain. This paper studies multimodal processing in the brain. Multimodal networks are aligned to the brain, and regions with better alignment are identified as the sites of multimodal processing. To verify that the alignment is actually due to multimodality, unimodal influence is removed from multimodal representations by subtracting out the multimodal target, as predicted by the unimodal input, from the multimodal features." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Relationship with previous work [1] and similar concurrent work [2,3] that uses multimodal models to probe for multimodal processing in the brain should be discussed. [1] studies cross-modal and jointly trained networks and uses a naturalistic multi-modal stimuli.\n- The random baseline described in 6.1 is good to make sure that the trained model weights matter. To better get a sense of whether the alignment actually reflects processing in the brain, another good sanity check would be to run a permutation test in which you keep the trained weights, but give the movie stimulus inputs to the model in scrambled order. This would give a floor for the scale of alignments that we see in subsequent results.\n\n## Small things\n- Line 276 typo: wmploy -> employ\n\n## References\n[1] Subramaniam, V., et al. \"Revealing Vision-Language Integration in the Brain with Multimodal Networks.\" International Conference on Machine Learning. International Conference on Machine Learning (ICML), 2024.\n\n[2] Kewenig, Viktor, et al. \"Evidence of Human-Like Visual-Linguistic Integration in Multimodal Large Language Models During Predictive Language Processing.\" arXiv preprint arXiv:2308.06035 (2023).\n\n[3] Dong, Dota Tianai, and Mariya Toneva. \"Vision-Language Integration in Multimodal Video Transformers (Partially) Aligns with the Brain.\" arXiv preprint arXiv:2311.07766 (2023)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "It would help to have additional methodological details in some sections:\n-\tWas cross-subject prediction generated by using all of the predictor subjects voxels, to predict the target subject’s voxel-wise responses?\n-\tWhat are the six different modalities in the image-bind modality? I thought it was an audio-visual model (which I would consider two modalities)\n-\tThe layer-wise predictions for models are shown in the appendix, but do the main text figures use all layers or best layer? If the latter, how is this layer selected.\n-\tAre the whole brain results averaged across all cortical voxels? Or only those with above-threshold cross-subject prediction? Or some other criteria?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "There are many interesting and novel aspects to this paper. First, while there have been extensive encoding model studies on visual or audio stimuli, few have looked at model comparison to multimodal movies. The comparison of audiovisual models to this data is particularly novel. Further the comparison of different types of multimodal models is interesting (though it is difficult to draw strong conclusions about what their comparison tells you about multimodal processing in the brain, see below), and particularly the attempts to quantify what additional explanatory power a multimodal model has over unimodal models. The encoding analyses were also robust, particularly the cross-movie train/test splits." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper compares different multimodal AI models to human brain responses while participants view audiovisual movies. They compare two different multimodal models, one cross-modal model that learns separate visual/audio embeddings and projects them into a shared representational space, and one jointly pretrained multimodal model, and three unimodal (vision or speech) models. The results show that in most brain regions multimodal training improves encoding model performance of voxel activity, particularly compared to unimodal speech models. The authors do additional residual analysis to understand the unique contribution of multimodal models (over unimodal models) to brain predictivity.\n\nOverall, this is an interesting approach and the paper has many strengths. The link between results and overall conclusions was not always clear. In particular, the small number of highly varied models makes it difficult to draw strong conclusions about parallels between multimodal processing in the models and brains. Finally, there were several clarification/presentation questions." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The biggest issues stem from the comparison of the performance of a relatively small number of models that differ along many factors. This limitation makes it difficult to attribute any model differences to multimodality or different cross-modal training schemes, as the model architecture and training sets vary from model to model. \n\nThe fMRI dataset uses a small-n, data-rich design, which is good, but given this, it would help to see the results at the individual subjects’ level (in the appendix). On bar plots, it would be nice to plot each of the six subjects as a point overlaid on the average bar (rather than error bars which can obscure differences across the small number of subjects).\n\nThe residual analyses and results are somewhat confusing. The residual correlation with unimodal features was 0.56, which is still quite high. Given this, it is not clear that unimodal information was removed from multimodal models. Alternatively, the authors could do the residual analysis on the brain instead of the models (i.e., fit a model with both unimodal and cross modal and predict with just cross modal). Relatedly they could calculate the product measure or unique variance explained by multimodal models above unimodal models from this joint model. \n\nOverall, the language and visual region responses look largely the same in most analyses. There are some quantitative/statistical differences, but the pattern is extremely similar. The authors should address this.\n\nPerhaps related to the above point, all regions of interest are defined anatomically, but there is a fair amount of subject-wise variability in the anatomy of high-level functional regions, such as language. The authors should address this limitation.\n\nAcronyms and plots were difficult to follow. Would help to spell out throughout vision versus lang regions for example, and clarify the legend in figure 4 (e.g., it was not clear on first read that “-“ indicates “minus”).\n\nFigure 5 is difficult to read and interpret. In terms of clarification, the authors should specify hemispheres/views (it looks like a lateral view on top, medial on bottom, but I’m not certain). The results look extremely noisy and seem to show random patterns, with as much red as blue. Blue are areas the unimodal models perform better? How should that be interpreted? The legend says the colorbar indicates “percentage increase/decrease. Does this mean 0.5% or 50%? If the former, these are very tiny differences, which perhaps explains the above confusion, but I believe makes it difficult to draw any strong conclusions about these results. \n\nI had questions about two of the conclusions listed in the discussion. I was unsure what the second part of conclusion 2 (“This variance can be partially attributed to the removal of video features alone, rather than auditory features.”) meant. I am also unconvinced of conclusion #4 given the overall similarity between language and visual brain regions described above.\n\nTypos:\n-\tLine 224: “audo” --> “audio”\n-\tLine 276: “wmploy” --> “employ”" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Why did the authors choose to use parametric statistics? To my knowledge non-parametric statistics are more common in NeuroAI to estimate a baseline performance rather than assuming one. \n\nAre the brains on the bottom in Figure 5 the medial view? It is difficult to see why there are four brains in each box. Outside of labeling, showing the sulci on the inflated brains would help to orient the reader to what is being shown." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "For the most part, the authors use standard and well validated methods to answer their question. I particularly like the approach of evaluating the performance on entire held-out videos and using the residual analysis to investing multi-modal effects above and beyond unimodal effects." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors aim to address whether there is a difference in brain-alignment based on whether multimodal models had cross-modality pretraining (separate encoders for two modalities) or joint pretraining (shared encoder across modalities). They compare one model of each type to video and speech models and evaluate the prediction in a large-scale open access movie dataset. Using residual analyses, they investigate whether there are multi-modal representations in visual and language regions of the brain." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The authors only evaluate two multi-modal models and a small number of unimodal models. However, the chosen models differ on many factors (e.g. architecture, training data) in addition to the input modality and as a result, it is premature to draw conclusions about semantic representations in the brain that may be attributable to any of these factors. To my mind, there are two ways to mitigate this concern: 1) controlled model training and evaluation so that only one factor varies at a time, or 2) testing many different models of a given class such that even across significant model variations there is a robust effect of modality regardless of particular model factors. I think that this is a serious concern because not all prior work has found that multi-modal models are more brain aligned. In controlled comparisons between visual models with the same architecture and training data, there was no performance increase as a result of contrastive image-language training (Conwell et al., 2023). These authors suggest that the higher alignment of CLIP relative to unimodal models in other work may be training set size. \n\nConwell, C., Prince, J. S., Kay, K. N., Alvarez, G. A., & Konkle, T. (2023). What can 1.8 billion regressions tell us about the pressures shaping high-level visual representation in brains and machines? (p. 2022.03.28.485868). bioRxiv. https://doi.org/10.1101/2022.03.28.485868\n\nA minor weakness of the paper is that the authors use custom, non-standard acronyms and names for brain regions (e.g., scene visual area, SV, and object visual processing region, OV). It is confusing as a reader, but more critically, it difficult to understand what has been found for particular regions across the field, making the status of the literature more tenuous. I would suggest that the authors adopt standard acronyms throughout (e.g., PPA instead of SV). \n\nAlthough the paper overall is fairly clear, section 6.3 and the corresponding figures (4, 9, and 10) are difficult to follow. I welcome clarification because, outside of a few lines in the discussion, I am having a hard time understanding which regions do show a multi-modal effect. Additionally, I think that the authors should emphasize whether they uncover expected unimodal effects in primary sensory cortices. In particular, we would expect that EVC would be predicted by visual models with no additional multi-modal contribution and similarly in AC but for auditory models. I am having trouble determining whether that is the case from the figures, but if it is not the case, it would lend more weight to my major concern about differences between models beyond modality." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024multimodal,\ntitle={Multi-modal brain encoding models for multi-modal stimuli},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=0dELcFHig2},\nnote={under review}\n}" }, "abstract": { "value": "Despite participants engaging in unimodal stimuli, such as watching images or silent videos, recent work has demonstrated that multi-modal Transformer models can predict visual brain activity impressively well, even with incongruent modality representations. This raises the question of how accurately these multi-modal models can predict brain activity when participants are engaged in multi-modal stimuli. As these models grow increasingly popular, their use in studying neural activity provides insights into how our brains respond to such multi-modal naturalistic stimuli, i.e., where it separates and integrates information across modalities through a hierarchy of early sensory regions to higher cognition (language regions). We investigate this question by using multiple unimodal and two types of multi-modal models—cross-modal and jointly pretrained—to determine which type of models is more relevant to fMRI brain activity when participants are engaged in watching movies (videos with audio). We observe that both types of multi-modal models show improved alignment in several language and visual regions. This study also helps in identifying which brain regions process unimodal versus multi-modal information. We further investigate the contribution of each modality to multi-modal alignment by carefully removing unimodal features one by one from multi-modal representations, and find that there is additional information beyond the unimodal embeddings that is processed in the visual and language regions. Based on this investigation, we find that while for cross-modal models, their brain alignment is partially attributed to the video modality; for jointly pretrained models, it is partially attributed to both the video and audio modalities. These findings serve as strong motivation for the neuro-science community to investigate the interpretability of these models for deepening our understanding of multi-modal information processing in brain." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "brain encoding", "fMRI", "multi-modal models", "multi-modal stimuli", "Transformers", "videos", "speech", "language" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/3f4478206dfab03a2644be4c503226053835f10a.pdf" }, "presentation": null, "primary_area": { "value": "applications to neuroscience & cognitive science" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Multi-modal brain encoding models for multi-modal stimuli" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. The layer propagation rule shows strong similarity to EGNN's Lower-bounded Residual Connection [1]. The paper needs to better elaborate on the key differences between these approaches.\n\n2. While EGNN appears as a baseline in Table 2, it is missing from the comprehensive comparison in Table 1. This makes it difficult to fully assess CDE-GNN's performance against this closely related method.\n\n3. Figure 1 analyzes the Dirichlet energy and edge space length for GAT, but lacks a corresponding visualization showing how CDE-GNN's Dirichlet energy behaves across different layers. Adding this visualization would help demonstrate the effectiveness of the proposed method in preventing energy decay.\n\n[1] Zhou et al. Dirichlet energy constrained learning for deep graph neural networks. Advances in Neural Information Processing Systems, 34, 2021." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. Well-structured presentation progressing from problem motivation to theoretical analysis to practical solution.\n2. Comprehensive empirical validation across multiple datasets." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper addresses the problem of over-smoothing in deep GNNs, where node embeddings become indistinguishable as network depth increases, leading to degraded performance. The authors present a novel geometric perspective on this issue and propose a method called Charge Dirichlet Energy (CDE-GNN). The authors validate their method through comprehensive experiments across various datasets and network depths, showing consistent performance improvements over baseline models." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Limited Novelty: using Dirichlet energy to overcome oversmoothness has been extensively studied.\n2. lack of detailed analysis of computational overhead compared to baseline methods" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. How does the model perform on large graph datasets, such as ogbn-products and ogbn-proteins?\n\n2. Can the authors provide validation on whether the over-smoothing problem is indeed addressed in practice on the experimental datasets?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The paper is well motivated and studies an important problem from an interesting perspective\n\n2. The paper is generally well written and easy to follow" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper studies the over-smoothing problem of deep graph neural networks and proposes a geometric perspective for addressing over-smoothing. Specifically, the authors analyze the Dirichlet energy minimized by the feed-forward computation process of GCNs and propose a new method based on Dirichlet energy for resolving over-smoothing when the layer number increases. Experiments on small datasets demonstrate the effectiveness of the model." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The proposed method has limited novelty given existing works that have explored similar ideas and model designs, e.g., [1,2]. Adding self-loop or residual link or strengthening the information of the centered nodes have been extensively used by existing GNN models, such as the well-known ones [1, 2] \n\n2. The theoretical results are not new and have been derived in the literature, e.g., [3, 4]. The result of Lemma 1 has been proved in [3] and [4]. Besides, the analysis presented in this paper only shows the result that is already well-known, i.e., over-smoothing will happen when the layer increases. There lacks analysis in why and how the propose model can address the over-smoothing.\n\n3. The experimental evaluation is limited in small datasets and comparison with state-of-the-arts is insufficient. More experiments on large datasets such as ogbn-products and ogbn-proteins are suggested. More comparison with state-of-the-art GNNs, especially the ones that can overcome over-smoothing, e.g., GCNII, are needed.\n\n[1] Simple and Deep Graph Convolutional Networks, ICML 2020\n\n[2] Predict then Propagate: Graph Neural Networks meet Personalized PageRank, ICLR 2019\n\n[4] A note on over-smoothing for graph neural networks. Arxiv 2020\n\n[5] Dirichlet Energy Constrained Learning for Deep Graph Neural Networks, NeurIPS 2022" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. How Eq. (8) could help preserve Dirichlet energy of each layer? By simply multiplying $X^{(l)}$ with a scalar and adding that to before applying nonlinearity, how would it guarantee Dirichlet energy is not vanishing?\n\n2. How is the approach compared with others like GCNII in terms of preserving DIrichlet energy? Intuitively adding initial residuals can already effectively preserve Dirichlet energy. Why the proposed approach adds the additional term before the nonlinearity and why the embedding of the previous layer instead of the initial layer is used?\n\n3. Why is the performance of all models on ogbn-arxiv much lower than officially reported results?\n\n4. Are there any plots of the layer-wise Dirichlet energy of the proposed model as well as some baselines (e.g., GCNII [1]) on these benchmarks? How does Dirichlet energy connect to actual performance? This would be important to help readers gain more intuition and also help understand the efficacy of the proposed approach.\n\n5. What is the rationale of introducing edge space (Definition 3.2) and how does it play a role in justifying the proposed method?\n\n\n[1] Chen et al. Simple and Deep Graph Convolutional Networks. ICML'20." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The problem tackled is important and this paper approaches it through the concept of Dirichlet energy.\n\n2. The method seems to offer empirical enhancements on various benchmarks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper discusses an approach to alleviate over-smoothing of deep graph neural networks through the lens of Dirichlet energy. The idea lies in adding one additional term in the layer-wise propagation which takes into account the Dirichlet energy of the initial graph. Experiments have been conducted on several node classification benchmarks showing that the model can yield better performance with increasing depth of the graph neural network." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Though a lot of efforts have been paid on discussing DIrichlet energy, how the proposed approach is linked to preserving Dirichlet energy still remains very unclear. Eq. (8) was introduced alone while more theoretical investigations and experimental observations should be incorporated.\n\n2. The presentation needs improvement. Wordy sentences present at times with many of them constantly repeated, e.g., the contents in Sec. 4.1 has been discussed multiple times in the previous sections and should be simplified for more informative contents.\n\n3. Some concepts were introduced with confusion and did not exhibit strong correlation to the proposed approach. For instance, how Proposition 3.2 is related to Eq. (8) (e.g., how Eq. (8) help address the vanishing problem) is unclear. There is also no clear reason of introducing the edge space (Definition 3.2) and Corollary 3.1 conveys limited information.\n\n4. Experiment settings are not convincing. For example, the reported performance of the baselines are remarkably lower than the official leaderboard of obgn-arxiv (https://ogb.stanford.edu/docs/leader_nodeprop/#ogbn-arxiv).\n\nMinor:\n\nMisuse of citet vs citep in multiple places hinders the readability. The authors are encouraged to correct these presentation issues.\n\nOverall, the paper in its current shape is unsatisfactory in justifying the rationale of the proposed approach, which is simply Eq. (8), and relevant discussions in both theory and experiment are missing, making it less self-consistent." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "1. In Equation (5), why summing over the edges twice?\n\n2. In Line 229-230, should Equation 9 actually refer to Equation 7?\n\n3. In Line 080, the authors state that the energy lower bound is learnable. However, as in Section 4.2.1, it is fixed as the initial energy. How is it learnable?\n\n4. In Line 304, it says the initial energy is multiplied by the initial embeddings, whereas in Equation 8 it is multiplied by the embeddings per layer. Which one is correct?\n\n5. In Table 1, why are there two bolded results for Citeseer and Film? Also, are those results for the semi-supervised or fully-supervised setting? It says in Line 334 and 370 that both settings’ results are in Table 1 but there are no marks for different settings.\n\n6. Several results in Table 2 and 3 don’t match. For example, on the Physics dataset with 32 layers, the optimal accuracy is 94.2 and 94.4, respectively. Why?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The paper proposed a very intuitive and generally applicable idea to effectively alleviate the over-smoothing problem for GNNs.\n2. The proposed idea is well-motivated by solid theoretical insights and results.\n3. The paper is easy to read." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Inspired by previous work on over-smoothing and Dirichlet energy, the authors propose a simple, intuitive, and generally applicable method named CDE-GNN to address the over-smoothing problem in Graph Neural Networks. The proposed approach is validated by theoretical and empirical results." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Some theoretical terms in Section 3 are not sufficiently introduced and explained. The authors are suggested to provide more detailed background and solid definition of the quantities involved.\n\n2. Several parts are repetitive or even inconsistent. For example, Section 4.1 is redundant and repetitive of the earlier content, as well as the Hyperparameter Analysis with its three following paragraphs in Section 5.2. Please refer to the Questions for details.\n\n3. The hyperparameter analysis is not insightful. The study of different activation functions, hidden dimensions, and dropout rates is old-fashioned and not unique to the proposed approach." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "**General Concerns**\n\n- Is $\\alpha$ in Equation(8) a learnable parameter or a manually defined hyper-parameter? The authors did not clarify this.\n- What are the exact values of $\\alpha$ and $E_{init}$ for each dataset in the experiments? These details are not mentioned in either the main text or the supplementary materials.\n\n- If $\\alpha$ is a hyper-parameter, the model's performance under different $\\alpha$ settings should be reported. Additionally, a discussion on how to select an appropriate $\\alpha$ would provide valuable insights for the community.\n- Why not include a comparison with EGNN in Table 1?\n- Why not use SReLU as the activation function in Section 5.2, given that it has been demonstrated in EGNN[1] for preserving Dirichlet energy, especially on large datasets like OGBN-arxiv?\n- It would be helpful if the authors could clarify the source of the baseline model performances in Table 1. Specifically, did they conduct all the experiments themselves, or were some of the results sourced from other papers?\n- Given that the benchmark results are based on 10 random splits, would it be possible to provide the standard deviation in addition to the mean? This could offer a more comprehensive understanding of the results.\n- It would be better to include publicly available code to ensure reproducibility;\n\n\\[1\\] K. Zhou et al., “Dirichlet energy constrained learning for deep graph neural networks,” in Advances in Neural Information Processing Systems, M. Ranzato, A. Beygelzimer, Y. Dauphin, P. S. Liang, and J. W. Vaughan, Eds., Curran Associates, Inc., 2021, pp. 21834–21846.\n\n\\[2\\] J. Zhu, Y. Yan, L. Zhao, M. Heimann, L. Akoglu, and D. Koutra, “Beyond Homophily in Graph Neural Networks: Current Limitations and Effective Designs,” in *Advances in Neural Information Processing Systems*, H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, and H. Lin, Eds., Curran Associates, Inc., 2020, pp. 7793–7804. \n\n\n**Ambiguous statement**\n\n- In Eqaution(8), $E_{init}$ is **multiplied by $X^{(l)}$** . While in Line 304, it states **multiplied by $X^{(0)}$** .\n\n > Line 304\n > The initial Dirichlet energy Einit captures the geometric information of the original graph and, when multiplied by the initial node embeddings X(0), ensures that each layer’s update process retains the topological features of the original graph.\n\n- data split question. What are the actual splits used for the Cora, Citeseer, and PubMed datasets? If Yang's split was adopted, why not use the same split as Geom-GCN for consistency?\n > Line 326\n >\n > For this study, we use the **Cora, Citeseer, and Pubmed** datasets Sen et al. (2008), **following the standard training/validation/test splits established by Yang et al. (2016)**\n >\n > Line 367\n >\n > We apply our model to datasets including **Cora, Citeseer, Pubmed**, Chameleon Rozemberczki et al.(2021), Film, Cornell, Texas, and Wisconsin, **following consistent splits of 48%, 32%, and 20%** for training, validation, and testing, respectively.\n\n\n**Minor comments**\n\n- Table1, Table 2, Table 3 out of page width;\n- In the introduction section, authors use the notion `a minimum Dirichlet energy ω` . But in the following text, $\\omega$ is no longer used; instead, $E_{init}$ is used. A consistent notation across the whole paper would be better;\n- There is a mistake in Equation (5): $\\mathcal{E}(f)$ is already a summation over the (i,j) pairs and cannot be summed again." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The authors revealed that the Dirichlet energy decay in deep GNNs is linearly proportional to the edge space sum with a constant $c$.\n\nThe authors provided that, within the Dirichlet energy analystic framework, it is crucial for the activation function $\\phi$ to satisfy $\\phi(0)=0$ in Proposition 3.2. This may indirectly explains why adding a shift $b$ in SReLU is effective.\n\nThe minimum Dirichlet energy constrained message updating scheme showed a good performance.\nExtensive benchmark performance comparisons." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a new understanding of Dirichlet energy in the context of GNNs, revealing the relationship between Dirichlet energy decay and edge space collapse. \n\nThe paper also introduces a new message updating scheme, which prevents over-smoothing by incorporating a residual term weighted by the minimum Dirichlet energy. And the authors conducted extensive comparison experiments demonstrating the effectiveness." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "**Lack of novelty**\n\nThis paper is a direct follow-up of [1].\n\nConsider equation(8) in [1] : \n\n$X^{(k)}=\\sigma\\left(\\left[\\left(1-c_{\\min }\\right) \\tilde{P} X^{(k-1)}+\\alpha X^{(k-1)}+\\beta X^{(0)}\\right] W^{(k)}\\right)$,\n\nwhere $\\alpha+\\beta=c_{\\text{min}}$.\n\nWhen set $\\beta=0$ and rephrase the equation as \n\n$X^{(k)}=\\sigma\\left(\\left[ \\tilde{P} X^{(k-1)}+ \\frac{c_{\\text{min}}}{\\left(1-c_{\\min }\\right)} X^{(k-1)} \\right]W^{(k)}\\right)$.\n\nReplacement the symbol $\\frac{1}{1-c_{\\text{min}}} \\to \\alpha$ and $c_{\\text{min}} \\to E_{init}$ , we immediately obtain:\n\n$X^{(k+1)}=\\sigma( \\tilde{P} X^{(k)}W^{(k)} + \\alpha E_{init} X^{(k)}W^{(k)}))$.\n\nCompare to Equation(8) in this paper:\n\n$X^{(l+1)}=\\sigma\\left(\\tilde{\\mathbf{L}} X^{(l)} \\mathbf{W}^{(l)}+\\alpha E_{\\text {init }} X^{(l)}\\right)$\n\nThese two update functions are remarkably similar, except for a weight matrix.\n\nIt has been demonstrated in [1] that, utilizing two distinct forms of residual connections can effectively constrain the lower bound of the DIRICHLET energy, and the constraint's intensity can be modulated by a gating parameter.\n\nThe approach presented in this paper appears to gracefully fit within this previously established framework, representing a specific instance when $\\beta=0$.\n\nThe main contribution is addressing the issue that researchers do not know how to choose an appropriate lower bound $c_{\\text{min}}$ for different datasets. While in this paper, the authors suggest that simply using the initial DIRICHLET energy $E_{\\text{init}}$ as the lower bound works very well.\n\nThe authors need to clarify how their method distinguishes itself from or enhances the previous approach.\nIt would be more valuable if the authors could elaborate on the significance of omitting the weight matrix in the update function and the rationale behind selecting the initial DIRICHLET energy as the lower bound among various initial value choices.\n\n[1] K. Zhou *et al.*, “Dirichlet energy constrained learning for deep graph neural networks,” in *Advances in Neural Information Processing Systems*, M. Ranzato, A. Beygelzimer, Y. Dauphin, P. S. Liang, and J. W. Vaughan, Eds., Curran Associates, Inc., 2021, pp. 21834–21846. \n\n \n\n**Lack of experiment**\n\n- Dirichelet energy visualization. Plotting the Dirichlet energy of each layer with and without the $E_{init}$ term would be more persuasive. As stated in Section 4.2, 'The initial Dirichlet energy serves as a lower bound for the Dirichlet energy,' it is expected that $E_{\\text{Dirichlet}}$ will converge to the lower bound $E_{init}$ as the number of layers increases.\n\n- The claim that $E_{init}$ preventing topological collapse in Section 4.2 should be supported by experimental evidence. Visualizing the node representations in the final layer and comparing them to the initial topology would be helpful. Consider using commonly employed techniques, such as t-SNE visualization\\[2\\]\\[3\\] or a color-propagation test[4].\n\n\n\n\n\\[2\\]D. Shen, C. Qin, Q. Zhang, H. Zhu, and H. Xiong, “Handling over-smoothing and over-squashing in graph convolution with maximization operation,” *IEEE Trans. Neural Netw. Learn. Syst.*, pp. 1–14, 2024, doi: [10.1109/TNNLS.2024.3442270](https://doi.org/10.1109/TNNLS.2024.3442270).\n\n\\[3\\]M. Liu, H. Gao, and S. Ji, “Towards Deeper Graph Neural Networks,” in *KDD*, Aug. 2020, pp. 338–348. doi: [10.1145/3394486.3403076](https://doi.org/10.1145/3394486.3403076).\n\n\\[4\\]K. Xu, C. Li, Y. Tian, T. Sonobe, K. Kawarabayashi, and S. Jegelka, “Representation learning on graphs with jumping knowledge networks,” in *International conference on machine learning*, PMLR, 2018, pp. 5453–5462." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024charge,\ntitle={{CHARGE} {DIRICHLET} {ENERGY}: Geometric Perspectives on Over-smoothing in Deep Graph Neural Networks},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=0e26yMOCbd},\nnote={under review}\n}" }, "abstract": { "value": "Over-smoothing is regarded as a key issue affecting the performance of deep Graph Neural Networks (GNNs). As the number of GNN layers increases, model performance degrades significantly, due to node embeddings converging into indistinguishable vectors. This phenomenon stems from the recursive aggregation of neighbor node representations, which impairs the distinguishability of node embeddings. From an energy perspective, this is associated with the convergence of node embeddings to a fixed point solution during the minimization of Dirichlet energy, hindering the model's ability to learn underlying geometric structures. While Graph Convolutional Networks (GCNs) have achieved success in modeling graph-structured data, there is still insufficient understanding of how the underlying geometry contributes to the trainability of deep GCNs.\nIn this paper, we present a novel geometric perspective to understand the poor performance of deep GCNs during training, a method called Charge Dirichlet Energy (\\model). We argue that maintaining a healthy geometric structure can significantly enhance the trainability of GCNs and enable state-of-the-art performance, even in base GCN architectures. Subsequently, we analyze the importance and feasibility of learning geometric shapes, demonstrating the critical role of geometric information in training deep GNNs. Extensive empirical validation on multiple benchmark datasets shows that our method improves the geometric shape of deep base GCNs, significantly enhancing their performance and outperforming many state-of-the-art methods in competitive settings. Our contributions include not only a new approach to mitigating over-smoothing and over-compression but also comprehensive theoretical and empirical verification of the importance of geometric structures for the trainability of deep GNNs." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Graph Neural Network", "Over-smoothing", "Dirichlet energy" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/ff9451b67025a24d1df2e8799f44a91ee720c765.pdf" }, "presentation": null, "primary_area": { "value": "learning on graphs and other geometries & topologies" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/c72f46c6226ed60e3749a3d2f277be122db8d0e4.pdf" }, "title": { "value": "CHARGE DIRICHLET ENERGY: Geometric Perspectives on Over-smoothing in Deep Graph Neural Networks" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "I have no ethical concerns on this work." }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "Q1. In contrast to scenarios where $\\pi_1$ dominates, how would the results in Figure 2 be affected if $\\pi_{opt}$ were dominant, a condition under which existing methods are known to perform well? Results for PN-GAIL in the Pendulum-v1 task with various $\\pi_{opt}:\\pi_1$ ratios are presented in Figure 7 of the supplementary material, but a more systematic comparison between PN-GAIL and baseline methods could strengthen the manuscript. If PN-GAIL demonstrates robust performance and outperforms the baseline methods in such scenarios, it would confirm the method's reliability. This evidence would support the assertion that PN-GAIL performs consistently well across a range of dataset quality distributions, thus setting it apart from existing methods.\n\nQ2. Instead of modifying the $\\pi_{opt}:\\pi_1$ ratio while maintaining a fixed total number of demonstrations, what would occur if the optimal demonstrations was kept invariant while the number of suboptimal demonstrations was varied? This setup would illustrate how PN-GAIL effectively utilizes additional suboptimal demonstrations, isolating the influence of optimal demonstrations on imitation performance. It would offer valuable insights into PN-GAIL’s ability to adapt to and leverage diverse demonstration qualities effectively." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "This work is well-motivated and effectively addresses the challenges presented in the previous study using sound methods.\n\nA notable strength of this work is its practice-oriented design of the objective functions, which enhances applicability in real-world scenarios.\nThis study removes the dependence on $\\eta$, the class prior $p(y=0)$ for imperfect demonstration datasets, from the primary objective. \nSince $\\eta$ is generally unknown and challenging to estimate, prior work treated it as a hyperparameter, requiring practitioners to invest substantial effort in tuning it. \nBy eliminating this reliance, the proposed approach reduces the overhead associated with hyperparameter optimization.\n\nAdditionally, the authors introduce near-optimal and practical choices for the parameters $\\alpha$ and $\\beta$ in the Balanced Semi-Confidence (BSC) objective, which can be straightforwardly calculated based on the dataset sizes $n_c$ (confidence-labeled) and $n_u$ (unlabeled). \nThis adjustment simplifies the implementation process and supports the broader applicability of imitation learning with imperfect demonstrations in practical settings.\n\nFurthermore, the manuscript includes theoretical analyses showing that (i) the proposed objective helps avoid the imitation of non-optimal data and (ii) derives a sample complexity bound for the BSC method, providing a rigorous foundation for the proposed improvements." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work addresses imitation learning from imperfect demonstrations, utilizing both confidence-labeled noisy demonstrations and unlabeled noisy demonstrations. It aims to overcome two primary limitations of prior work, specifically 2IWIL [A], a representative approach to this problem that employs a two-step learning process: (i) semi-confidence labeler training on the unlabeled dataset, and (ii) confidence-based generative imitation learning.\n\nThe proposed method, PN-GAIL (Positive-Negative GAIL), tackles the limitations of 2IWIL as follows:\n\n1. **Incorporating Negative Risk to Objective**: 2IWIL overlooks the negative risk associated with imperfect demonstrations, leading the discriminator to disproportionately prioritize the positive risk of frequent samples. PN-GAIL addresses this by incorporating both positive and negative risks into the confidence-based imitation learning objective, ensuring a more reliable evaluation regarding to demonstration quality.\n2. **Balanced Semi-Confidence (BSC) Classification**: In 2IWIL, semi-confidence (SC) classification is used to train a confidence labeler for unlabeled demonstrations. However, SC classification tends to overestimate the confidence of labeled data and underestimate the confidence of unlabeled data. To address this, PN-GAIL introduces a balanced semi-confidence (BSC) objective and further suggests near-optimal values for hyperparameters $\\alpha$ and $\\beta$, enhancing practical applicability.\n\n[A] Wu et al., \"Imitation Learning from Imperfect Demonstration,\" ICML 2019." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Despite the many strengths of this work, the empirical results presented in the manuscript do not stand out as particularly impressive.\n\nSpecifically, Figure 1 shows that the performance difference between PN-GAIL and the most competitive baseline across tasks is not significant. In my opinion, as discussed in Section 2, since baseline methods typically assume a dominant proportion of $\\pi_{opt}$, exploring scenarios with a more skewed ratio (e.g., $\\pi_{opt}:\\pi_1=1:10$) might provide a more notable results where conventional methods fail while PN-GAIL successes. \nI think conducting experiments with more extreme demonstration ratios could more clearly demonstrate the scenarios in which PN-GAIL offers a distinct advantage over baseline methods.\n\n**[Minor Comments]**\n1. For Figures 2, 3, and 4, using distinct line colors for different methods would enhance readability.\n2. In the ablation study presented in Figure 3, it would be advantageous to include results from 2IWIL—even though they are already provided in Figure 2. Since 2IWIL represents a variant of PN-GAIL that excludes both the PN objective and the BSC, its direct comparison within the same figure would clarify the incremental benefits of these components." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "* Theorem 1’s bound relies on knowing the variances. However, the variances can be difficult to estimate in real-world applications. Given this dependency, how practically useful is the bound in scenarios where variance values are unknown or hard to assess?\n* Theorem 2’s bound may become quite loose if the Rademacher complexity is high, as is typical with deep and wide neural networks. Could this have negative implications for the method's reliability when using complex models?\n* Related to my disagree on the claim that 'GAIL treats non-optimal demonstrations as if they were optimal' in the weaknesses section, please could you provide a counterargument?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "* Originality: The use of positive and negative risks to manage imperfect demos tackles some of the real-world problems of non-optimal data in a practical way.\n* Quality: The paper provides theoretical analysis for the positive-negative risk approach and shows experimental results across multiple control tasks.\n* Clarity: Most theoretical ideas are clearly presented with good notation.\n* Significance: The approach is relevant for real-world applications. It might have a real impact on the usability of IL in practical scenarios." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces an IL method that handles imperfect demos by including both the optimal and non-optimal data in training, assuming that the demos come with a confidence score. The goal is to improve policy learning when demos aren’t perfectly optimal. The method assigns weights to optimal and non-optimal examples through a semi-supervised confidence classifier. Experiments on six control tasks show that PN-GAIL performs better than standard GAIL and other baselines under these imperfect conditions." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* The method’s reliance on confidence scores may limit its applicability when it’s difficult to assign confidence levels directly. Human annotation of preferences over trajectories, for instance, might be more feasible than assigning explicit confidence scores.\n* The fundamental motivation of this paper is based on a claim that 'GAIL treats non-optimal demonstrations as if they were optimal'. I disagree with this claim. GAIL’s objective is to minimize the JS divergence between the expert and agent trajectory distributions, aiming to reproduce the overall distribution of expert demonstrations. When the agent policy is close to the expert policy, the discriminator's output tends to be 0.5 everywhere. If expert demos include a mix of optimal and sub-optimal trajectories, GAIL should naturally capture this mixture without necessarily assuming optimality. Could you provide a counterargument or clarification on why PN-GAIL’s approach is necessary, given this perspective?\n* Lacks comparison with other advanced IL algorithms, such as f-IRL." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. Could the authors highlight expert performance more prominently in the figures to enhance clarity and interpretability?\n\n2. Would varying values of $n_{u}$ and $n_{c}$ significantly impact the performance of the proposed method? Additionally, could the authors provide guidelines or criteria for selecting optimal values for these parameters in practice?\n\n3. Figure 1 offers a valuable comparison of the proposed method against GAIL and 2IWIL. To further support this comparison, it would be better if the authors could also provide more intuition of why GAIL and 2IWIL fail but the proposed method succeed. E.g., gradient colors for different confidence scores, and why 2IWIL fails to predict them. Additionally, Why GAIL predicts 5.0 for some of theses data points?\n\n4. To further contextualize this work within imitation learning (specifically, driver behavior imitation), it would be beneficial to incorporate additional relevant studies, such as:\n\n[2] Ruan, Kangrui, and Xuan Di. \"Learning human driving behaviors with sequential causal imitation learning.\" Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 36. No. 4. 2022.\n\n[3] Hawke, Jeffrey, et al. \"Urban driving with conditional imitation learning.\" 2020 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2020.\n\nand so on." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "This paper introduces a new algorithm (based on 2IWIL and IC-GAIL) supported by theoretical analysis and demonstrates its effectiveness across multiple tasks. Experiments are extensive, including different benchmarks, different $\\pi_{OPT} : \\pi_{1}$ ratios, different standard deviations of Gaussian noise." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors propose PN-GAIL, an extension of the GAIL framework designed to handle the imperfect expert demonstrations. By predicting confidence scores for unlabeled data, PN-GAIL allows for more accurate imitation learning without relying solely on optimal examples. Theoretical analysis is provided for the output of the optimal discriminator in the proposed method." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Although the proposed method is based on [1], e.g., some techniques to prove Theorem 4.1, Theorem 4.2, have been explored in [1], this study sill broadens the scope of [1]. One potential weakness is: although $\\alpha$ and $\\beta$ are intended to play distinct roles in Theorem 4.2, they are selected to be identical in Algorithm 1, which may affect the overall applicability of the algorithm. \n\n[1] Wu, Yueh-Hua, et al. \"Imitation learning from imperfect demonstration.\" International Conference on Machine Learning. PMLR, 2019." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "1. About clarification: PN-GAIL\\BSC: PN-GAIL without balanced semi-conf (BSC) classification--does this mean no classification used or SC used? Without a probabilistic classifier, how to obtain confidence scores? \n\n2. To verify BSC outperform SC in 2IWIL, straightforward comparisons are PN-GAIL(BSC) vs. PN-GAIL(switch to SC) vs. 2IWIL(switch to BSC) vs. 2IWIL(SC). Why do you consider the way of comparisons that are presented in the paper?\n\n3. It is better to provide more detailed explanations and discussions regarding those figures. For example, the statement about Figure 3 is quite short. The audience would like to learn about what possible reasons are that result in the varying performance patterns across six environments in Figure 3. We can observe that in some cases three colors (methods) are similar; in some cases blue and green are close; in some cases, blue and orange are close; while in some cases, blue works best. Could you provide a systematic analysis?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "1. The main objective/purpose of this paper is articulated explicitly with the existing methods and their limitations highlighted clearly.\n2. The literature review on IL, GAIL, and IL with imperfect information is comprehensive. The preliminaries provide the essential information of 2IWIL.\n3. The theoretical derivations regarding the discriminator modification and classification refinement are straightforward and concise in the main body, which makes audience easy to follow; meanwhile, the supplemental materials in the appendices provide necessary, detailed explanations.\n4. The experiments with three goals setup are tightly related to the main achievements that the paper wants to claim. The experiments are conducted with representative benchmarks across various environments, effectively showing the performances with well-formatted figures and table in the main context.\n5. Overall, the authors identify the potential challenges of 2IWIL about certain non-optimal data with high frequencies in an unlabeled demonstration set significantly affecting reward and policy generation. The topic is likely to be of interest to a large proportion of the community. The proposed PN-GAIL creatively update the previous methods to successfully remove limitations of prior IL results to a certain extent." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Motivated by the limitation of 2IWIL in a type of scenario where certain non-optimal demonstrations have high probabilities of appearing in a set of imperfect (unlabeled) demonstrations, the paper proposes a new method named PN-GAIL, better leveraging optimal and non-optimal information from imperfect demonstrations to learn optimal reward and policy by assessing both positive and negative risks. Moreover, the paper modifies semi-conf classification in 2IWIL to establish balanced semi-conf classification to handle better the cases where certain demonstrations only exist in the confidence (known, labeled) data.\n\nThe authors conduct experiments, comparing their PN-GAIL to four baseline methods across six environments. The results show that PN-GAIL alleviates the impact of the unbalanced frequency in impact demonstrations, outperforms other methods, and maintains relatively good performances given the decreasing number of labels. Also, the outcomes demonstrate that the balanced semi-conf classifier improves performances, particularly in three out of six environments." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Missing definition: in Section 3, what δ represents. Please define δ when it is first introduced in the paper.\n\n2. Needing clarification: PN-GAIL\\BSC: PN-GAIL without balanced semi-conf (BSC) classification--does this mean no classification used or SC used? Without a probabilistic classifier, how to obtain confidence scores? Please explicitly state whether a SC classification is used, and to explain how confidence scores are obtained if no classifier is used. \n\n3. Lack of analysis for experimental outcomes: it is necessary to provide more detailed explanations and discussions regarding those figures. For example, in Figure 3 what possible reasons (e.g., characteristics of each environment? limited number of demonstrations?) are that result in the varying performance patterns across six environments. We can observe that in some cases three colors (methods) are similar; in some cases blue and green are close; in some cases, blue and orange are close; while in some cases, blue works best. Please provide a systematic analysis of how different factors might contributing to these patterns." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We introduce a Positive-Negative Generative Adversarial Imitation Learning (PN-GAIL) method within the framework of Generative Adversarial Imitation Learning (GAIL) to leverage non-optimal information from imperfect demonstrations." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024pngail,\ntitle={{PN}-{GAIL}: Leveraging Non-optimal Information from Imperfect Demonstrations},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=0e2pcSxQJS},\nnote={under review}\n}" }, "abstract": { "value": "Imitation learning aims at constructing an optimal policy by emulating expert demonstrations. However, the prevailing approaches in this domain typically presume that the demonstrations are optimal, an assumption that seldom holds true in the complexities of real-world applications. The data collected in practical scenarios often contains imperfections, encompassing both optimal and non-optimal examples. In this study, we propose Positive-Negative Generative Adversarial Imitation Learning (PN-GAIL), a novel approach that falls within the framework of Generative Adversarial Imitation Learning (GAIL). PN-GAIL innovatively leverages non-optimal information from imperfect demonstrations, allowing the discriminator to comprehensively assess the positive and negative risks associated with these demonstrations. Furthermore, it requires only a small subset of labeled confidence scores. Theoretical analysis indicates that PN-GAIL deviates from the non-optimal data while mimicking imperfect demonstrations. Experimental results demonstrate that PN-GAIL surpasses conventional baseline methods in dealing with imperfect demonstrations, thereby significantly augmenting the practical utility of imitation learning in real-world contexts. Our codes are available at https://anonymous.4open.science/r/PN-GAIL-3828." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Generative adversarial imitation learning", "imperfect demonstrations", "reinforcement learning" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/42aad028d937cdabcffb5638874918e590128c5b.pdf" }, "presentation": null, "primary_area": { "value": "reinforcement learning" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "PN-GAIL: Leveraging Non-optimal Information from Imperfect Demonstrations" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1、 more diverse set of low-resource languages in the experimental dataset will be helpful\n2、 the impact of various auxiliary languages can be deeply analyzed\n3、 prompt analyzation can be improved" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. Interesting research, Introduces Mufu, a novel approach leveraging multilingual context and post-editing for low-resource language translation.\n2. Employs automatically generated candidates and instructions to correct translations, enhancing LLM's reasoning capability.\n3. Demonstrates robustness against poor-quality auxiliary translations, outperforming specialized NMT systems in many low-resource pairs.\n4. Proposes a hybrid learning paradigm, combining in-context learning and finetuning for improved translation quality.\n5. Implements knowledge distillation to reduce inference costs while maintaining performance gains in low-resource translations." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces \"Mufu\" , which is a method for low-resource language translation using a multilingual fused learning approach, specifically targeting large language models (LLMs).\nThe Mufu method, which aims to address the challenge that large language models (LLMs) perform well in translating high-resource languages but still struggle with low-resource languages. The Mufu prompting approach turns the translation task into a post-editing task, leveraging the reasoning capabilities of LLMs with auxiliary translation candidates, requiring the model to assess input quality, align semantics cross-lingually, copy from relevant inputs, and override incorrect instances. Experiments show that LLMs fine-tuned with Mufu-style prompts achieve better performance than the NLLB 1.3B distilled model in 64% of low- and very-low-resource language pairs on the Flores-200 dataset." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Experiment Method Optimization， Consider incorporating a more diverse set of low-resource languages in the experimental dataset to better generalize the findings and evaluate the model's performance across a wider linguistic spectrum.\n\n2. Experiment Conclusion Enhancement， Suggest conducting ablation studies to isolate the specific contributions of different components of Mufu, such as the impact of various auxiliary languages, to fine-tune the approach and maximize translation accuracy.\n\n3. 5-shot Prompting Improvement， Explore the use of meta-learning strategies in 5-shot prompting to enhance the model's ability to quickly adapt to new translation tasks with limited examples, potentially improving the efficiency of the learning process." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Please see the weaknesses for the questions." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- experimental results show some effectiveness of the proposed approach\n- the idea of leveraging multilinguality via the prompt sounds technically good" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper tackles low-resource translation quality improvement in LLM models. To maximize data efficiency in the low-resource setting, the authors introduce a new approach called Mufu, including automatic selection of multilingual translation candidates and an instruction tuning to correct inaccurate translations via the prompt. Experimental results on Flores-200 dataset for English-XX directions show robustness and achieves better performance against NLLB 1.3B distilled model in 64% of low- and very-low resource language pairs." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- unclear about the experimental results; how to decide the best prompt template for mufu; any impacts of language combination used in the prompt template - for example, have you ever tried adding high-resource language translation pairs during training to enhance multilingual training with high and low-resource language pairs?\n- results are not convincing enough, maybe due to low-resource setting with limited improvement in ChrF. Can you report other metrics such as sacreBLEU scores? Have you tried finetuning LLM with low-resource monolingual data so that the LLM can more effectively enhance Mufu." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "- Were there any accidental translations in a different language for Mufu{5,10,20}?\n- What exactly is Win% vs teacher? For instance, for NLLB 1.3B distilled, its chrF is 46.0 whereas that of teacher is 43.7, still its win% is 41.3? It means NLLB 1.3B was less than 50% correct when compared to teacher model still its chrF score is higher? Another example, Win% vs teacher is 56.2 for NLLB 54B MoE (48.9 chrF) whereas for mufulora20 with PaLM2 S it is 99% with chrF less than NLLB 54B MoE on FLORES 200. It will be great if authors can formalise what is Win% vs teacher.\n- Can the authors explain In theory… model outputs (line 207-211)?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The paper is very clearly written and easy to follow.\n- They combine 2 interesting learning paradigms: ICL and parameter tuning and their core focus is on very-low and low resource language which I really liked. \n- They perform evaluation on NTREX which is important for ood evaluation.\n- The experiments performed by authors are quite extensive. I especially liked mufu5hrl, mufu5tr, distilled, and lora which corroborate their approach of selecting 5,10,20 related languages from URIEL.\n- Quantitative evidence provided in Figure 3 is quite helpful in knowing how language transfer is taking place. Moreover, the attention pattern further helps in understanding how attention pattern is making mufu models perform better." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces Mufu, which turns translation into post-editing task by providing auxiliary translations and target translation from teacher model. The student model learns in-context to produce the correct target translation and is then fine-tuned against references. Languages for auxiliary translations are chosen from URIEL and they evaluate using PaLM S family models along with Gemma 2B,7B on FLORES 200 (iid), and NTREX (ood). The paper contains thorough ablation studies as well as cross lingual attention alignment which helps understanding or interpreting how model is learning through in-context." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- No model sizes available for PaLM2 family of models. I’m not sure how to compare them with Gemma or NLLB.\n- If I were to just compare on the basis of chrF score, only PaLM2 XXS -NLT and PaLM2 S are able to beat NLLB 1.3B distilled model in both FLORES 200 and NTREX (and Gemma 7B on FLORES 200). Rest all are inferior to NLLB 1.3B distilled. One suggestion for authors in this case will be to add `Latency` column for all models (higher for mufu and lower for distilled models) to show the trade off between accuracy and latency which will help readers understand how competitive other models are.\n- The authors have mentioned this but finetuning an LLM (or even NLLB with 1B+ param) with just 787 sentences and in-context learning will definitely lead to overfitting which is evident by the fact that mufu20lora performed better than full finetuning. I wonder if that is the case for other models too? \n- It’s great they used Gemma 2, an open weight model but I’m slightly disappointed that majority of their experiments use PaLM2 models which are not public like Gemma 2. \n- Two iteration process (teacher model followed by student model) is quite expensive. The authors have mentioned that distillation helps to alleviate the problem but it only worked for NTREX in PaLM2 XXS - NTL (not for Gemma 7B), performance on FLORES 200 for both distilled models is lower than NLLB 1.3B. \n- The authors experiment with one learning paradigm i.e., in-context learning for LLMs for distillation. Did they try distillation from model outputs (not the one fine-tuned with mufu20)? How much better or worse is in-context learning compared to vanilla distillation?" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We finetune LLMs for translation by prompting with multilingual context in order to harness their cross-lingual reasoning, and substantially improve the models' performance in low-resource translation." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024mufu,\ntitle={Mufu: Multilingual Fused Learning for Low-Resource Translation with {LLM}},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=0eMsrRMmCw},\nnote={under review}\n}" }, "abstract": { "value": "Multilingual large language models (LLMs) are great translators, but this is largely limited to high-resource languages. For many LLMs, translating in and out of low-resource languages remains a challenging task. To maximize data efficiency in this low-resource setting, we introduce Mufu, which includes a selection of automatically generated multilingual candidates and an instruction to correct inaccurate translations in the prompt. Mufu prompts turn a translation task into a postediting one, and seek to harness the LLM’s reasoning capability with auxiliary translation candidates, from which the model is required to assess the input quality, align the semantics cross-lingually, copy from relevant inputs and override instances that are incorrect. Our experiments on En-XX translations over the Flores-200 dataset show LLMs finetuned against Mufu-style prompts are robust to poor quality auxiliary translation candidates, achieving performance superior to NLLB 1.3B distilled model in 64% of low- and very-low-resource language pairs. We then distill these models to reduce inference cost, while maintaining on average 3.1 chrF improvement over finetune-only baseline in low-resource translations." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "translation", "low-resource", "large language model" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/84c5f101f608bcec3450bc726c1edb69e63752a8.pdf" }, "presentation": null, "primary_area": { "value": "applications to computer vision, audio, language, and other modalities" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Mufu: Multilingual Fused Learning for Low-Resource Translation with LLM" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please refer to the 'weaknesses' part." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1.\tThe paper is well-written and easy understanding. The figure 1 and 6 are intuitive to understand the overall framework.\n2.\tThe saliency technique used for analyzing the information flow among various tokens are interesting and intuitive. The conclusion that visual information injection dominates in shallow layers while intra-visual aggregation dominates in deeper layers makes sense.\n3.\tExperimental results demonstrate the effectiveness of the proposed method to some extent." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper dive into how multimodal large language models (MLLMs) process and utilize visual information. Based on the widely used saliency technique for interpretability, information flow among different tokens across different layers is analyzed. The authors find out that visual information injection dominates in shallow layers while intra-visual aggregation dominates in deeper layers. Finally, hierarchical image token pruning is proposed to prune at both shallow and deep layer with specific criterion." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.\tThe phenomenon analyzed in the paper is not surprising, and previous works [1][2] have pointed out the similar findings that information of vision tokens has migrated to following text tokens within the first few layers of MLLMs. Thus, I think this paper is with limited novelty which employs a commonly used technique to analyze the phenomenon that have been identified.\n2.\tI wonder how the parameter K1, K2 is determined. For different datasets and tasks, the parameters may be different. Directly setting k1 and k2 to a pre-defined value may be not suitable. Could K1, K2 be dynamically adjusted based on the input samples?\n3.\tThe evaluation datasets used in the paper is quite limited. I suggest the authors to evaluate on other commonly used datasets, especially OCR-related or fine-grained datasets to demonstrate the effectiveness, e.g., textvqa, gqa, docvqa, chartqa, seed-bench. For the efficiency evaluation, I suggest the authors to include inference time and GPU memory.\n4.\tTwo different criteria are used in shallow and deep layers. I wonder the performance if the same criterion is used. If the performance is similar, the analysis of different information flow in shallow and deep layers is not very convincing. \n\n[1] An Image is Worth 1/2 Tokens After Layer 2: Plug-and-PLay Acceleration for VLLM Inference.\n\n[2] BOOSTING MULTIMODAL LARGE LANGUAGE MODELS WITH VISUAL TOKENS WITHDRAWAL FOR RAPID INFERENCE" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "I noticed that VTM suggests that visual tokens are not essential in the deeper layers of MLLMs and strategically withdraw them at a certain layer. I'm very curious about the advantages of HiMAP compared to VTM." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The paper is well-written, with clear explanations and helpful visuals.\n\n2. This paper introduces intriguing hypotheses and includes extensive, detailed studies to support them.\n\n3. Extensive experiments confirm HiMAP’s effectiveness, showing reduced computational costs while preserving performance." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper identifies the minimal role of image tokens in MLLM predictions, uncovers patterns in visual-textual interactions, introduces HiMAP as a pruning technique to reduce inference latency without compromising performance, and demonstrates its effectiveness across diverse vision-language tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Lack of ablation studies. For example, additional experiments could be included to examine the impact of various pruning strategies on the model and to assess the effects of different hyperparameter settings, such as K1, K2 and the ratio.\n\n2. The authors might consider including additional benchmarks, such as MME and AI2D, and presenting fine-grained performance scores. Additionally, it would be helpful to include metrics such as GPU memory and total time in the comparisons to provide a more comprehensive evaluation." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. As mentioned in Weakness #1.\n2. In Section 2.2, the authors derive two main factors based on insights 1) \"As the model depth exceeds...\" and 2) \"Instruction tokens exert the most...\". The first conclusion is undoubtedly correct, but the second remains questionable. Although much work has validated the redundancy of image tokens in MLLMs, the two insights provided in this paper do not directly lead to this conclusion. The \"limited impact of image tokens\" mentioned in the paper only supports the first conclusion, while the argument for the second conclusion would require a computation of saliency for each image token (assuming a length of N), and if the authors conducted this experiment, they would find that only some image tokens have high significance." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "1. Universality: As a universal image token pruning algorithm, HiMAP can be easily applied to different architectures of MLLMs to achieve accelerated inference.\n2. Usability: The method is straightforward, with low transfer costs.\n3. The saliency scores and dynamic pruning approach used in the paper can provide inspiration for the field of accelerated inference." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper examines the significance of image tokens in different layers of MLLMs, suggesting that image tokens tend to facilitate modality injection in shallow layers and engage in more internal interactions in deeper layers. Based on this analysis, the paper proposes HiMAP, an algorithm for dynamically pruning image tokens to accelerate inference, which has been validated for its effectiveness across various multimodal tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The paper lacks line numbers, which seems to deviate from ICLR's submission standards and may hinder reviewers in accurately pinpointing issues within the document.\n2. The experiments are not comprehensive enough, with validation only on a limited number of tasks. As a universal solution, it should be tested on common multimodal benchmarks, such as LLava-Bench, MMBench, etc." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "* Does the type of task potentially influence the conclusion regarding the minimal importance of visual tokens, which account for only 0.03% of the significance of textual tokens? For example, the proportion of visual tokens may considerably decrease in multi-turn dialogue tasks. At the same time, their relative significance could increase due to the normalization of sequence length reflected in Equations 2, 3, and 4. \n* The insight presented in line 7 on page 4 looks weird. If the contribution of tokens from deeper layers to response prediction is low, why not leverage tokens from the shallow layer with the most significant contribution to generating responses? In this reviewer's opinion, only the comparison within each layer in Figure 2 carries practical significance.\n* The reviewer suggests evaluating the performance of HiMAP against the baseline on fine-grained perception tasks, such as document understanding and OCR (e.g., Chartqa[1] and Docvqa[2]). This would provide a more solid demonstration of HiMAP's efficacy in reducing redundant image tokens. \n\n[1] Chartqa: A benchmark for question answering about charts with visual and logical reasoning. \\\n[2] Docvqa: A dataset for vqa on document images." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "* From the novel perspective of information flow, the authors have analyzed the fusion patterns of visual features in MLLMs. Building upon this analysis, they proposed an adaptive approach for visual redundancy elimination.\n* The structural design of HiMAP is intuitive and demonstrates robust performance across a range of tasks, exemplified by image captioning and VQA.\n* The article is highly readable, featuring a well-defined and clear structure." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes that in MLLMs, image tokens primarily convey visual information to instruction tokens in shallow layers, while deeper layers consolidate the remaining visual data. Based on this insight, a plug-and-play visual pruning method, HiMAP, is proposed to reduce the computational costs." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* In analyzing the information flow between visual tokens and textual tokens, it is essential to thoroughly examine the flow in both directions. Hypothesis H1 is valid only if it is determined that the primary flow of information occurs from the visual modality to the textual modality, rather than in the opposite direction. This requires conducting an experiment to compare the magnitudes of $S_{vt}$ in Equation 6 with $S_{tv}$.\n\n* Furthermore, merely showing a decline in performance by restricting the interaction between image tokens and instruction tokens in shallow layers does not sufficiently support Hypothesis H1. It is essential to complement this with an experiment that specifically limits the flow of intra-visual information within shallow layers (the current IMG2RND experiment in Figure 4 is not direct). Only when the resulting performance degradation is considerably less pronounced can Hypothesis H1 be adequately substantiated." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "Please refer to weakness." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1) The paper is well motivated, authors design the HiMAP strategy corresponding to the behavior of how visual information is utilized in MLLM layers.\n\n2) HiMAP manages to reduce the computational costs by approximately 65% without sacrificing performance." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this paper, authors begin by studying how visual tokens interact in MLLMs and observe that 1) image tokens interact strongly with text instruction tokens to form cross-modal representations in shallow layers; 2) image tokens aggregate remaining visual information in deeper layers. Upon this, they propose a token pruning inference strategy, HiMAP, for MLLM, by selecting the most important image tokens by image-text-attention scores in shallow layers and image-self-attention scores in deeper layers." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1) The ablation studies are insufficient, e.g., different choices of K1, K2 , R1, R2; ablation for using different importance criteria on shallow layers and deep layers. \n\n2) The finding that LLMs may well process visual tokens in the early layers has already been proposed in previous works[1-3]. The stagewise token pruning strategy has also been proposed for efficient MLLM [3]. Consequently, the novelty of this paper is somewhat limited.\n\n3) More benchmarks for MLLM performance evaluation should be involved to exhibit the effectiveness of HiMAP, e.g., the widely used GQA, MME, MM-Vet, VQAv2 and etc.\nThe paper should be drafted with line number on the page and the writing of the paper should be improved.\n\n\n[1] An Image is Worth 1/2 Tokens After Layer 2: Plug-and-PLay Acceleration for VLLM Inference, in ECCV24.\n\n[2] DeepStack: Deeply Stacking Visual Tokens is Surprisingly Simple and Effective for LMMs, in NeurIPS24.\n\n[3] LLaVolta: Efficient Multi-modal Models via Stage-wise Visual Context Compression, in NeurIPS24." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "This paper uncovers a shift in visual information processing in MLLMs and introduces a novel image token pruning method to accelerate inference." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024unraveling,\ntitle={Unraveling the Shift of Visual Information Flow in {MLLM}s: From Phased Interaction to Efficient Inference},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=0eRJRbVG95},\nnote={under review}\n}" }, "abstract": { "value": "Multimodal large language models (MLLMs) improve performance on vision-language tasks by integrating visual features from pre-trained vision encoders into large language models (LLMs). However, how MLLMs process and utilize visual information remains unclear. In this paper, a shift in the dominant flow of visual information is uncovered: (1) in shallow layers, strong interactions are observed between image tokens and instruction tokens, where most visual information is injected into instruction tokens to form cross-modal semantic representations; (2) in deeper layers, image tokens primarily interact with each other, aggregating the remaining visual information to optimize semantic representations within the visual modality. Based on these insights, we propose Hierarchical Modality-Aware Pruning (HiMAP), a plug-and-play inference acceleration method that dynamically prunes image tokens at specific layers, reducing computational costs by approximately 65% without sacrificing performance. Our findings offer a new understanding of visual information processing in MLLMs and provide a state-of-the-art solution for efficient inference. Code is released at https://anonymous.4open.science/r/HiMAP." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Multimodal Large Language Models", "Visual Information Flow", "Inference Acceleration" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/b921e97813be3d37a1eaed29a01d6160817ed83f.pdf" }, "presentation": null, "primary_area": { "value": "interpretability and explainable AI" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Unraveling the Shift of Visual Information Flow in MLLMs: From Phased Interaction to Efficient Inference" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "* RewardMath is based on the dataset MATH500, where does the dataset MATH500 come from? Is MATH500 prior work (and if yes the citation is missing) or is this a contribution of the paper (in this case it should be made clear).\n* Does MATH500 address the incorrect annotation problem found in PRM800K? \n* Can authors also show evaluations and ablations on gsm8k [1] and MATH [2] which are the most common eval tasks for LLM reasoning capabilities?\n* Authors identify that in RewardBench the accepted response often has less steps than rejected ones, which could give a chance for models to reward hack (i.e. reward relies on the number of steps instead of the actual response quality). Did the authors ablate this? I.e. Does the reward-hacking model predict lower reward if we make the accepted response in RewardMath longer? And vice versa, does it predict higher reward if we make the rejected response shorter? \n* Regarding RewardMath giving more than one rejected responses: If one is trying to do preference learning using a llama model as the base model, is it important for the reward model to know the rejected response generated by a non-llama model should be worse than the accepted response? I.e. the distribution could be very different that it never encounters it during preference learning. I.e. for Figure 4, if we use a llama model as the policy, does one-to-many RewardMath still do better than one-to-one RewardMath chosen & Llama rejection?\n* Does reward-model free alignment methods like DPO also suffer from reward model overfitting problem? What is the advantage of using reward models over reward model-free methods for reasoning tasks? \n* Does the benchmark evaluate cases where both the rejected and chosen response arrive at the same answer, but the rejected answer has the wrong steps? I.e. this is common for truth or false questions. \n\n\n[1] Training Verifiers to Solve Math Word Problems\n\n[2] Measuring mathematical problem solving with the math dataset" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "* The topic on how to Improve LLM reasoning capabilities has recently gained a lot of attention. This paper focuses on having good benchmarks for evaluating these efforts, and this could be very impactful if done correctly. \n\n* Authors identify flaws of existing benchmarks and make good efforts to fix them. \n\n* Paper has good results, specifically Figure 4 is very cool showing RewardMath has stronger correlation with downstream tasks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Authors aims to design a better benchmark for evaluating reward models in reasoning tasks. Authors identify problems with the previous benchmark RewardBench and proposes RewardMath:\n* RewardBench is based on PRM800K which contains many wrong annotations, RewardMath instead is based on MATH500.\n* RewardBench uses pairs of (human annotated example, LLM annotated incorrect example). RewardMath includes more than one incorrect example.\n* RewardBench’s accepted and rejected examples have different number of steps, this could be a spurious correlation that leads to reward hacking. RewardMath fixes this.\n* RewardBench’s PRM evals uses product instead of mean which biases shorter responses, authors fix this by using mean instead. \n\nTo demonstrate the improvements of rewardmath 1) authors compare performances of different reward models on both RewardBench and RewardMath, and show rewardmath has higher correlation with downstream evals 2) authors show rewardmath exhibits the common trend of larger dataset -> better performance while rewardbench does not." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "See questions I have below" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- As the proxy reward model trained from synthetic data shows far from optimal performance in Table 3 (only around 13% for RewardMATH and 69% for RewardBench), can you consider using better proxy RMs? The increase in value from 12.68 to 13.51 is not very convincing to me that this is a strong trend.\n\n- In Figure 6, the gold reward (or even the oracle reward) does not drop for most cases even with the maximum KL distance considered. If a larger N is considered for BoN sampling, will the graph drop down as in Gao et al [2]? For a larger N is RewardMATH still successful in detecting more robust reward models regarding the overoptimization problem?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The paper provides clear and sufficient empirical evidence that their RewardMATH benchmark is more reliable than the math subset of RewardBench [1]. The empirical results are also clear as LLM policy using BoN on high-scored reward models on RewardBench shows little to no correlation with the performance increase of Math benchmarks (r-square = 0-0.1), while RewardMATH shows a much stronger correlation \n(r-square = 0.6-0.8) in Figure 3.\n- The authors have evaluated diverse reward models on RewardMATH, including LLMs (generative reward models), classifier-based reward models, and process reward models.\n- The paper considers the problem of over-optimization using a synthetic setup of gold RMs and proxy RMs. \n\n[1] https://arxiv.org/abs/2403.13787" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes RewardMATH, a reward model evaluation benchmark focused on the math domain. This benchmark adopts a one-to-many comparison to evaluate reward models robustly. Specifically, it provides 10 responses for each prompt where only 1 response is the answer. The evaluated reward model is considered accurate only when the answer response is given the highest score among all 10 responses. \nThe authors also empirically provide sufficient evidence that the RewardMATH benchmark provides reliable estimates for reward models that give policies optimized using BoN are indeed robust in math benchmarks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The work would be more interesting if the authors showed any other domains (such as coding or text summarisation or maybe safety) reward model benchmark can be improved by the framework proposed here (by adopting multiple responses and using diverse LLMs to generate outputs). Any initial or limited experiments would be helpful. \n\n- The lack of PPO (or DPO) usage for policy fine-tuning in experiments seems like a major weakness. The main contribution of this paper is using policy fine-tuning methods to verify if the RewardMATH benchmark scores correlate with the signals it provides during policy fine-tuning. I agree with this approach and am impressed by the number of experiments conducted to verify this using mainly Best-of-N sampling. However, Best-of-N sampling is an inference time method to generate better model outputs using reward models, whereas PPO (or possibly DPO) is the main fine-tuning method researchers use. Although Figure 5 does show a PPO experiment under a synthetic setup, the number of checkpoints or whether the dots follow the findings from Gao et al [2] is not clear to me. Without any solid PPO results, Best of N sampling seems not enough to verify the benchmark's capability of measuring the robustness of reward models. The work will be much more convincing if the authors show more PPO-trained policy evaluations. Or at least, it will be helpful if the author provides more context as to why PPO is hard to train in their non-synthetic setup. Also, I suspect high-scoring reward models on RewardMATH have the ability to find the best response from multiple responses, and Best-of-N adopts a very similar way as it picks the response with the highest reward, resulting in a high correlation of results. Whether this ability will generalize even on PPO setups is not clear to me at this point. \n\n- Experiment results in Figure 6 compare diverse RMs on both RewardBench and RewardMATH benchmarks with gold or oracle rewards. It would be nice if the authors not only provided the numbers but also a statistical analysis (such as Kendell's tau) that measures the agreement between RewardMATH(or RewardBench) and oracle (or gold) reward scores in Figure 6. \n\n[2] https://arxiv.org/abs/2210.10760" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. [Clarification] Are the prompts used to evaluate the LLM judge on REWARDMATH the same as the prompt used to evaluate the LLM judge on the Reward Bench? Different prompting strategy (eg. difference system prompt) raises concerns regarding fair comparison between the two benchmarks.\n2. [Clarification] What is MATH500? The author did not mention the details behind this dataset which they used for their benchmark. Were there any steps taken to ensure the dataset is not contaminated with the models being evaluated? If the dataset is used during training on any of the evaluated RMs, the benchmark’s reliability will be undermined. \n3. [Clarification] What was the motivation behind different parts of the Synthetic Data experiment? What was the reasoning behind using the MetaMATH dataset? Why was only 80K out of the 155K data points augmented from MATH used for training? \n4. The authors did not mention how the incorrect solutions are ensured to be actually incorrect. Were there steps taken to validate the said incorrect solutions are indeed incorrect?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. **Thoroughness**: The paper presents detailed implementations, including training hyperparameters and experimental protocols. This ensures that other researchers can accurately reproduce the experiments and validate the findings. \n2. **Relevance**: This work addresses a critical gap in the field by focusing on reward model evaluation, a crucial area of research that has significant implications for the development of more reliable AI systems.\n3. **Motivation**: The paper presents a compelling critique of the existing Reward Bench evaluation metric, establishing a strong foundation for their work. The authors make a persuasive case for developing benchmarks that minimize over-optimization risks, backing their arguments with experimental evidence. This dual focus on improving metrics while addressing practical concerns demonstrates clear motivation for the research." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a new benchmark, REWARDMATH, to improve the robustness evaluation of reward models in mathematical reasoning tasks. It highlights limitations in the existing RewardBench benchmark, which relies on single comparisons between chosen and rejected solutions, potentially leading to reward hacking and misjudgments of model robustness. REWARDMATH addresses this by using one-to-many comparisons with multiple incorrect solutions to better capture robustness and reduce the risk of reward over-optimization. Experimental results indicate that scores on REWARDMATH strongly correlate with policy optimization outcomes and provide a more reliable measure of reward model robustness compared to RewardBench. This benchmark aims to enhance the development and reliability of RLHF systems, with the findings underscoring the potential of REWARDMATH to serve as a trustworthy evaluation tool in this domain​." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. **Clarity**: The paper is generally well written, however, it has some clarity issues, especially in section 5, which is hard to follow. Clarification questions are asked in the question section, marked with [Clarification]. The authors should address those questions. \n2. **Benchmark Biases**: The paper has several biases, raising concerns on the claimed robustness and reliability. Examples and comments below:\n\n> Line 206: Hence, we first convert the human-annotated solutions from MATH500 into step-by-step machine-generated solutions. We prompt GPT-4, using 4 carefully crafted exemplars for each math subject as part of the prompt.\n\nAll correct solutions in the benchmark are generated via GPT-4, raising concerns regarding biases towards GPT-series models. Even though the authors manually inspect the solutions, the solutions were still mainly generated using GPT-4. Notably, the authors observe LLM judges from the GPT-4 Series to perform significantly higher than other models (Line 286), which is likely due to this oversight (since it is known LLM judges tend to bias their own response, eg. GPT-4 family judge favors responses from GPT-4 family). The authors should use a diverse set of LLMs to curate the correct solutions to avoid potential biases.\n\n> Line 805: Secondly, we instruct GPT-4-0125-preview to select a specific step from the correct solution, transform it into an erroneous step, and then prompt again to continue generating the solutions from the erroneous step.\n\nSimilar to the previous point, employing GPT-4-0125-preview as editor to insert errors into other LLMs’ answers may introduce biases. Additional validation is needed to ensure the benchmark does not exhibit any bias towards GPT family models. \n\n> Line 402: We assume Internlm2-7B-reward, which performs well on both RewardBench and REWARDMATH, as the gold RM.\n\nThe use of Internlm2-7b-reward as the gold standard lacks sufficient justification and raises several concerns about experimental validity. The author relies primarily on performance metrics from RewardBench and REWARDMATH, but this approach is problematic for multiple reasons. First, the authors themselves criticized RewardBench for containing incorrect ground truth data and failing to adequately assess reward models. Second, using REWARDMATH as a benchmark is circular since it's the very dataset being studied. High scores on these benchmarks alone don't necessarily indicate that a reward model can reliably approximate human preferences. To establish Internlm2-7b-reward as a legitimate gold standard, the author should conduct additional validation studies specifically demonstrating its ability to align with human judgment on mathematical tasks.\n\n> Line 426: We find that proxy reward models trained on smaller datasets reach peak rewards at lower KL divergences, indicating faster over-optimization.\n\nThe author assumes KL divergence adequately captures optimization degree without proper justification. KL may not account for other important aspects of policy change. Further study will strengthen the experimental results. \n\n3. **Comprehensiveness**: This paper's scope is notably narrow, focusing solely on evaluating reward models' performance on mathematical problems. While it attempts to address limitations in a small subset of the Reward Bench dataset, its improvements remain constrained. The study primarily concentrates on reward over-optimization, overlooking other potential vulnerabilities in reward model benchmarking. Additionally, the benchmark's methodology of comparing one correct solution against multiple incorrect ones limits its thoroughness. Furthermore, the author's assumption that MATH500 adequately represents mathematical reasoning tasks may be oversimplified. These limitations collectively suggest a need for a more comprehensive approach to reward model evaluation." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Why do the authors choose 10 generations with 9 incorrect and 1 correct answer in the k-way comparisons? How does the choice of k and numbers of correct and incorrect answers affect the resulting correlations and reward overoptimization? The relationship between reward overoptimization and the proposed benchmark needs more rigorous analysis." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The paper provides a good correlation study between the proposed benchmark and best of N downstream performance in datasets like MATH and Gaokao etc. The proposed one-to-many comparison seems to be on the right direction for better correlation with downstream performance." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces REWARDMATH, a benchmark for evaluating reward models in mathematical reasoning tasks, arguing that it provides a more reliable evaluation than the existing RewardBench by using one-to-many comparisons instead of pairwise comparisons. The authors validate their approach by showing correlation between benchmark performance and best-of-n sampling results." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I have the following major concerns:\n\n1. Technical Contribution & Novelty: The primary contribution seems to be replacing pairwise comparisons in reward bench with best of N comparisons, which is an incremental modification rather than a substantial methodological advance. I do not think the change only is sufficient for a publication in top machine learning conference. The correlation between N-way comparison performance and best-of-N sampling results is somewhat expected and doesn't provide deep insights into reward model behavior. \n\n\n2. Unclear Definition of Robustness: The paper uses \"robustness\" throughout but never provides a precise definition. The authors seem to equate robustness with correlation to best-of-n sampling results, but this is circular reasoning since the benchmark itself uses similar methodology. There's no clear framework for what properties a \"robust\" reward model should have beyond empirical correlation with certain metrics. \n\n\n3. Limited Experimental Validation: The paper relies heavily on correlation with best-of-n sampling as validation, but doesn't explore other important aspects of reward model behavior. To make the paper deeper and broader, it would be great if the authors could try comparing the correlation of different real down-stream fine-tuning techniques like BoN SFT, DPO, PPO etc. and see whether how RewardBench and RewardMath correlate with downstream performance there. It would also be interesting to see if such observation extends to other domain like coding, and perhaps even open-ended generations without ground truth." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Q. How much of the relatively poor results for RewardBench is due to the noisy annotations inherited from PRM800K, as mentioned in Sec. 3.1? In other words, could simply fixing these annotations significantly change the comparison?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The paper identifies a particular limitation of RewardBench, a popular benchmark for evaluating reward models, in assessing mathematical reasoning and introduces a new benchmark that addresses this issue by including one-to-many comparison data.\n2. The authors provide extensive experiments and analyses across different reward model types, including both proprietary and open models, and assess various performance metrics." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper addresses a specific limitation of RewardBench, a widely used benchmark for reward model evaluation, in its assessment of mathematical reasoning capabilities. To this end, the authors introduce RewardMATH, a new benchmark that employs one-to-many comparisons of chosen and rejected responses to mathematical questions to enhance evaluation robustness. Experiments show that RewardMATH correlates well with policy performance and is more effective at identifying potential reward overoptimization and the reliability of reward signals." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The benchmark comparison primarily involves RewardBench, which is designed to evaluate reward models more holistically across various domains. However, is the comparison in terms of mathematical reasoning appropriate, given that RewardMATH is specifically designed for this purpose? If RewardBench is indeed the most comprehensive eval set even for mathematical reasoning tasks prior to RewardMATH, it would be helpful to clarify this point.\n2. The benchmark appears to lack cases where reward models must distinguish between correct solutions of varying quality, such as those missing reasoning steps. It is also unclear whether 500 samples is sufficient to cover diverse mathematical reasoning tasks.\n3. Tables 1 and 2 report performance comparisons of various LLMs on RewardBench and RewardMATH. The results seem to merely suggest that the two benchmarks differ significantly. Can we conclude from these results that \"high scores on RewardBench do not guarantee **robustness** in reward models\"?" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose a design for a reliable benchmark for reward models and validate our design using the results of optimized policies and through the lens of reward overoptimization." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024evaluating,\ntitle={Evaluating Robustness of Reward Models for Mathematical Reasoning},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=0er6aOyXUD},\nnote={under review}\n}" }, "abstract": { "value": "Reward models are key in reinforcement learning from human feedback (RLHF) systems, aligning the model behavior with human preferences.\nParticularly in the math domain, there have been plenty of studies using reward models to align policies for improving reasoning capabilities.\nRecently, as the importance of reward models has been emphasized, RewardBench is proposed to understand their behavior.\nHowever, we figure out that the math subset of RewardBench has different representations between chosen and rejected completions, and relies on a single comparison, which may lead to unreliable results as it only see an isolated case.\nTherefore, it fails to accurately present the robustness of reward models, leading to a misunderstanding of its performance and potentially resulting in reward hacking.\nIn this work, we introduce a new design for reliable evaluation of reward models, and to validate this, we construct RewardMATH, a benchmark that effectively represents the robustness of reward models in mathematical reasoning tasks.\nWe demonstrate that the scores on RewardMATH strongly correlate with the results of optimized policy and effectively estimate reward overoptimization, whereas the existing benchmark shows almost no correlation.\nThe results underscore the potential of our design to enhance the reliability of evaluation, and represent the robustness of reward model.\nWe make our code and data publicly available." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "mathematical reasoning", "RLHF", "reward models", "reward overoptimization", "language models", "benchmark" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/1d4b478bfde9d6708c1e37c9fa5ce32c5f4d1a17.pdf" }, "presentation": null, "primary_area": { "value": "datasets and benchmarks" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/f881b604bd9898f154a5ed11cc10b7f855da77d6.zip" }, "title": { "value": "Evaluating Robustness of Reward Models for Mathematical Reasoning" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- What are the specific challenges encountered when integrating CPSEM and RFA into existing architectures?\n- Have you considered applying AHR to non-vision data, such as text or audio?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "+ The combination of generative and exemplar replay in a single system that minimizes memory while maintaining high performance is novel.\n+ The proposed method achieves a significant reduction in memory requirements (O(0.1t)), which is crucial for scalability in CIL.\n+ Comprehensive experiments across multiple benchmarks and comparisons with state-of-the-art (SOTA) methods demonstrate the robustness of AHR." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents a novel approach to CIL called Autoencoder-based Hybrid Replay (AHR). This method combines exemplar and generative replay techniques to address key challenges in CIL, such as task confusion and catastrophic forgetting (CF). The hybrid autoencoder (HAE) serves as both a discriminative and generative model, storing data in a compressed latent space with minimal memory (O(0.1t)) compared to traditional exemplar methods (O(t)). The use of charged particle system energy minimization (CPSEM) and a repulsive force algorithm (RFA) aids in optimal placement of class centroids in the latent space. The experimental results indicate that AHR consistently outperforms existing baselines across five benchmarks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "-The paper lacks exploration of real-world applications or more complex, dynamic scenarios beyond standard benchmarks.\n-Performance could be impacted if the autoencoder's compression and reconstruction capabilities are not well-optimized.\n- While memory reduction is emphasized, the impact of this method on significantly larger-scale datasets or more diverse data distributions is not detailed." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 1 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "NA" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "I have read this paper carefully. Unfortunately, this paper is totally out of my research area. Therefore, I cannot capture the brilliance of this paper." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. This paper is well organized.\n\n2. I appreciate the extensive experiments.\n\n3. The idea of modeling the energy dynamics within the system akin to charged particles is interesting." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes an autoencoder-based hybrid replay (AHR) strategy that leverages our new hybrid autoencoder (HAE) to function as\na compressor to alleviate the requirement for large memory, achieving O(0.1t) at the worst case with the computing complexity of O(t) while accomplishing state-of-the-art performance." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The writing of Section 2 (\"OUR STRATEGY: AUTOENCODER-BASED HYBRID REPLAY (AHR)\") is confusing. Please outline the motivations of each step and explain why it makes sense.\n\n2. The technique contribution is fuzzy. I want to know what technique used in this paper and what technical challenge or a novel idea in the technique of proposed method." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "please refer to weakness section" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1) Easy to follow and addresses the important topic of computational burden for continual learning algorithms\n2) The motivation behind the Hybrid replay is well written by contextualizing the current works of literature along with their gaps" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces Hybrid Autoencoder (HAE) and Autoencoder-Based Hybrid Replay (AHR) strategies to reduce the memory burden for CIL, especially for replay-based approaches. HAE combines both discriminative and generative modeling to handle classification and replay tasks. It employs Charged Particle System Energy Minimization (CPSEM) equations and the Repulsive Force Algorithm (RFA) to manage class separation within its latent space, enabling class identification using Euclidean distance. AHR integrates exemplar and generative replay strategies by storing samples in the latent space, which significantly reduces memory usage. Its decoder is designed to memorize training data, allowing for effective replay without the typical issues of hazy pseudo-data found in other generative approaches. Simulations in various benchmark datasets also validate the hypothesis." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1) Inconsistent notation\n: In the explanation of AHR (184) you are referring are T_l but in the algorithm, it is T_i which is the same for P\n : what does the * refer to in algorithm 1?\n : What is J_l?\n : ^ in explanation and ' is used in Figure 1 are interchangeably used for generated output\n : what is \\mathcal{T} in Figure 1?\n 2) Is there any explanation for how the memory is reduced to 0.1t\n\n 3) It is claimed that the complexity reduces to 10%, but no empirical evidence is provided to validate that hypothesis\n4) Evaluation metric: It is unclear to readers, (line 373) does the accuracy represents the accuracy on the only last task after training on all tasks or is average on all previous tasks.\n5) How does the number of exemplars decrease over time? You are representing the exemplars in latent space? Does it mean reducing the number of classes?\n(Table 3) Why the different numbers of epochs for AHR and others? Please be clear on the size of latent and raw ? is 150 (latent) better than 20 (raw)\n6) It would be clearer to the readers if there was some explanation of how CPSEM and RFA create incremental embeddings\n7) The main objective of this paper is to reduce the size of exemplars in memory. In the related work section, the authors focus on mainly describing the current replay mechanism without mentioning how the current strategies fall short in reducing the size and its relation to AHR\n\n\n8) There is no comparative analysis of the work with state-of-the-art replay methods such as:\n I) Rolnick, David, et al. \"Experience replay for continual learning.\" Advances in neural information processing systems 32 (2019).\n\n II) Buzzega, Pietro, et al. \"Dark experience for general continual learning: a strong, simple baseline.\" Advances in neural information processing systems 33 (2020): 15920-15930.\n\nwhich makes it challenging to assess the significance of the work in the literature.\n\n### Comments on evaluation: \nI am mainly concerned about the accuracies for CIFAR100 and mimiImageNet. There are various works [FeTrIL [1] by Petit, Grégoire, et al., FeCAM [2] by Goswami et al] utilizing ResNet-18/32 achieving higher accuracy of more than 65% even in exemplar-free settings. I wonder how with an exemplar, the model is not able to maintain that accuracy.\n\n[1] Petit, Grégoire, et al. \"Fetril: Feature translation for exemplar-free class-incremental learning.\" Proceedings of the IEEE/CVF winter conference on applications of computer vision. 2023.\n\n[2] Goswami, Dipam, et al. \"Fecam: Exploiting the heterogeneity of class distributions in exemplar-free continual learning.\" Advances in Neural Information Processing Systems 36 (2024)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "Overall, I believe the paper has significant issues with presentation (as outlined in points [a-e] of the weakness section), which necessitate a major reformatting. Additionally, regarding other concerns, the novelty appears limited to performing classification at the encoder level and the introduction of CPSEM for initializing class centroids. While the latter is a novel contribution, it is neither well-explained nor well-motivated (as noted in points [f-g] of the weakness section). The experimental section, particularly the method comparison, needs refinement by considering recent related work on the proposed approach, utilizing higher-resolution benchmarks, and employing architectures with a larger latent space (as indicated in points [h-i] of the weakness section). Moreover, the details about the training resources should be clarified (point [l] in the weakness section).\n\nI believe the paper has the potential for significant improvement for a future submission. To enhance its novelty, the authors should focus more on the description and proposal of the class centroid initialization, which could provide both theoretical and empirical insights. However, as previously mentioned, these insights are lacking in the current version. Additionally, improving the experimental section would further strengthen the submission.\n\nConsidering all the above, I recommend rejecting the current submission. The paper is not yet ready for publication and requires significant revisions. I suggest making these improvements and submitting it to a different venue." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The Charged Particle System Energy Minimization (CPSEM) method for initializing class centroids for the encoder training is novel and interesting.\n- Several comparisons are carried out in the experimental section, and the method shows good performance on the benchmarks and methodologies used for comparison.\n- Detailed analysis on resource consumption is provided.\n- The ablation study comparing the proposed AHR with AHR using original images (AHR-lossless) highlights that the quality of the images generated by the encoder is sufficiently good for replay, as AHR with original images achieves similar performance. I appreciated this analysis." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper addresses the problem of Class-Incremental Learning by introducing a Hybrid Autoencoder (HAE). The proposed model is designed for both exemplar replay during incremental training and classification at inference. The autoencoder consists of two components: an encoder, which is trained to minimize the Euclidean distance between the latent representation and the corresponding class centroid, and a decoder, which is trained to minimize the reconstruction error between the input and output images. Both the encoder and decoder are trained with a distillation loss on the previous task to mitigate activation drift.\n\nAt the encoder level, the class centroids serve as anchors in the latent space, guiding the latent representations towards their respective classes. These centroids are initialized before training using the Charged Particle System Energy Minimization (CPSEM) method, which ensures that the centroids are well-separated. After training, a nearest-mean classification rule is applied to classify test images based on the proximity of their latent representations to these centroids.\n\nIn the post-training phase, the encoder's latent space output is used to populate a replay buffer with latent representations of the current task samples, following a herding strategy. These representations are replayed in the next task by feeding them into the decoder trained on the previous task, which generates the corresponding images. These generated images are then used for training on the new task. Instead of storing images, as is typically done in incremental learning, the proposed method reduces memory requirements by storing only the latent space representations.\n\nThe authors compare their approach to several incremental learning methods on benchmarks such as MNIST, Balanced SVHN, CIFAR-10, and miniImageNet. They also analyze the method's resource consumption, evaluate different decoder sizes, and provide an ablation study by comparing its performance when real images are used during the replay phase instead of the generated ones." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Overall, I believe the presentation of the paper requires significant improvement. Below are my major concerns regarding the presentation:\n\n- (a) The complexity analysis in the introduction needs to be clarified and expanded. The notation $O(0.1t)$ is incorrect according to the definition of Big-O notation. What does this represent? If the authors intend to convey that memory is saved by storing the latent space representation, I suggest incorporating both the latent space dimension and image size into the complexity analysis. Additionally, the term $e$ in $O(cte)$ is not defined and requires clarification.\n\n - (b) The notation throughout the paper is difficult to follow, with multiple indices used unnecessarily (e.g., in Equation 1). This excessive use of notation makes the paper hard to read. I suggest simplifying the notation wherever possible.\n\n- (c) Equation 1 seems to imply that all examples from previous tasks are needed to minimize the reconstruction error. However, I understand that this is not the case—some examples are real images, while others are generated. The replay buffer should be explicitly highlighted to clarify this in the equation.\n\n- (d) The three pseudocode blocks on page 4 make the methodology difficult to follow. Including all three pseudocode blocks on a single page compresses the accompanying methodology description into less than half a page. As a result, some LaTeX formulas in the main text break across two lines, further increasing the difficulty of reading.\n\n - (e) The organization of the paper should be reconsidered. Given the limited space for submission, dedicating more than two pages to the literature review while allocating just over one page to the methodology does not allow for a proper description of the proposed approach. I suggest moving the extended literature review to the appendix and presenting a more concise version in the main paper.\n\nRegarding the methodology and experimental section, my major concerns are as follows:\n\n- (f) The introduction of the Charged Particle System Energy Minimization (CPSEM) for initializing class centroids is interesting but requires additional explanation. It is unclear why this type of initialization benefits the autoencoder and how it relates to the Coulomb interaction energy . While I do not expect a full background on the calculus of variations for minimizing energy, more mathematical details—even in the appendix—would be helpful. An analysis of how centroids are distributed in the latent space is required to underline why the proposed strategy is effective. Furthermore, the CCE (class centroid embedding) placement is explained only through pseudocode, with no accompanying description. At a minimum, the operations performed in Algorithm 2 should be explained in words to provide intuition, especially for readers unfamiliar with the physics-based intuition behind this algorithm.\n\n- (g) Regarding CCE, why is this initialization considered effective? If the goal is to initialize centroids such that the class centroids are distant from each other, why not simply use the K-means algorithm? Alternatively, why not select class centroids as the latent vectors for each class that are the most distant from each other in terms of Euclidean distance, in a similar way as performed with hard negative sampling?\n\n- (h) The usage of the latent space for memory reduction and the decoder for latent replay is not novel. For example, Ayub et al. (ICLR 2021) [1] employ the encoder for storing latent representations and replay these latent representations in subsequent incremental learning steps. When the memory budget is reached, the latent representations are compressed into centroids and covariances. A comparison with their approach is necessary. The storage of latent representations for incremental learning and efficient memory replay with autoencoder is also explored in [2].\n\n- (i) Comparison in Table 2. 1) Some comparisons are unnecessary since the methods the authors compare to perform different tasks. For instance, Prediction Error Based Classification (PEC) [3] is designed for online continual learning (single-pass data), while the authors address the problem of offline incremental learning. It is clear that the harder setting for which PEC is designed results in lower performance compared to the authors' method. The authors should compare their approach only with offline class incremental learning methods under the same conditions. 2) Since the authors' method operates in an offline incremental learning setting, they should compare it with recent exemplar-based class incremental approaches, such as X-DER [4] and MEMO [5]. Additionally, the comparison should consider using a larger and more realistic backbone, such as ResNet-18, which is now commonly evaluated with more parameters and a larger latent space [4][5]. The paper should also evaluate how the method performs with higher-resolution images (e.g., 224x224) on a dataset like ImageNet100 [5].\n\n- (l) In Table 3, the authors report the wall-clock time for training their methods. They state that training takes about 8 hours on CIFAR-100 with ResNet-32, which seems excessive. In FACIL [6], joint training on a not particularly novel GPU requires less time, as only about 400k parameters need to be optimized. What are the timings for joint training with the same epoch budget? An incremental learning method should be more efficient than joint training. Additionally, the authors should specify the device used for the experiments when reporting training time.\n\n[1] A. Ayub and A. Wagner, “{EEC}: Learning to encode and regenerate images for continual learning,” in International\nConference on Learning Representations, 2021.\n\n[2] Caccia, L., Belilovsky, E., Caccia, M. &amp; Pineau, J.. (2020). Online Learned Continual Compression with Adaptive Quantization Modules. Proceedings of the 37th International Conference on Machine Learning, in Proceedings of Machine Learning Research.\n\n[3] Michał Zaj ˛ac, Tinne Tuytelaars, and Gido M van de Ven. Prediction error-based classification for\nclass-incremental learning, in ICLR 2024\n\n[4]Zhou, Da-Wei and Wang, Qi-Wei and Ye, Han-Jia and Zhan, De-Chuan, A Model or 603 Exemplars: Towards Memory-Efficient Class-Incremental Learning, in ICLR 2023 \n\n[5] Matteo Boschini, Lorenzo Bonicelli, Pietro Buzzega, Angelo Porrello, Simone Calderara. Class-Incremental Continual Learning into the eXtended DER-verse, in TPAMI 2022 \n\n\n[6] Marc Masana, Xialei Liu, Bartlomiej Twardowski, Mikel Menta, Andrew D. Bagdanov, Joost van de Weijer Class-incremental learning: survey and performance evaluation, TPAMI 2022" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "A memory-efficient architecture for incremental learning based on model-based and exemplar-based approach." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024autoencoderbased,\ntitle={Autoencoder-Based Hybrid Replay for Class-Incremental Learning},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=0eu837jdBD},\nnote={under review}\n}" }, "abstract": { "value": "In class-incremental learning (CIL), effective incremental learning strategies are essential to mitigate task confusion and catastrophic forgetting, especially as the number of tasks $t$ increases. Current exemplar replay strategies impose $\\mathcal{O}(t)$ memory/compute complexities. We propose an autoencoder-based hybrid replay (AHR) strategy that leverages our new hybrid autoencoder (HAE) to function as a compressor to alleviate the requirement for large memory, achieving $\\mathcal{O}(0.1 t)$ at the worst case with the computing complexity of $\\mathcal{O}(t)$ while accomplishing state-of-the-art performance. The decoder later recovers the exemplar data stored in the latent space, rather than in raw format. Additionally, HAE is designed for both discriminative and generative modeling, enabling classification and replay capabilities, respectively. HAE adopts the charged particle system energy minimization equations and repulsive force algorithm for the incremental embedding and distribution of new class centroids in its latent space. Our results demonstrate that AHR consistently outperforms recent baselines across multiple benchmarks while operating with the same memory/compute budgets." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Catastrophic Forgetting", "Class-Incremental Learning", "Continual Learning", "Task Confusion." ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/d5d18926236cda4e00b58d03c40e762b89ecc4c5.pdf" }, "presentation": null, "primary_area": { "value": "transfer learning, meta learning, and lifelong learning" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/1e7fc3634266b38a80c28780735cf0abb3833d13.zip" }, "title": { "value": "Autoencoder-Based Hybrid Replay for Class-Incremental Learning" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "While the analyses show some interesting trends, it is difficult to tell how meaningful or significant the numerical differences are. Methods for analyzing LLM layers other than through ID could have been discussed in a prior work section." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper conducts a broad analysis across 5 different LLMs and considers a range of questions and ablation studies (e.g., estimating ID on shuffled data, comparing layers across different models); altogether an impressively broad set of experiments. The paper is clearly written and presents a few new insights (e.g., correlation between peak onset and performance). Code and data would be made available, which would be valuable for the community." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper uses the technique of intrinsic dimension estimation as a tool for analyzing properties of different transformer LLM layers. 5 different LLMs are analyzed on textual inputs from 3 different public-domain corpora. In addition to computing the intrinsic dimensionality (ID) (using the generalized ratios intrinsic dimension estimator) for different layers, the ID is correlated with performance of different layers' representations on syntactic and semantic probing tasks. Furthermore, the difference in representational power between different layers is measured using an Information Imbalance criterion. The authors find that middle layers in LLMs have the highest ID; ID peaks seem to be an indicator of linguistic structure being learnt; early onset of peaks in ID across layers is correlated with better next token prediction performance performance; and high ID peak layers are representationally equivalent across different LLMs." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The use of ID as an analysis tool for LLM layers is not an entirely new idea (e.g., https://arxiv.org/pdf/2402.18048). \nMost of the results (e.g., the peaking of ID at middle layers, emergence of linguistically informative representations in those layers) has been shown before by means of other methods (e.g., mutual information or canonical correlation analysis). These should have been discussed in more detail under prior work." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "(1) Please check the questions listed under Weaknesses." }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "(1)\tAlthough inspired by the earlier work of (Valeriani et al., 2023), this work greatly extends the models investigated to 5 distinct mainstay transformer-decoder-only LLMs and added more extensive probing and downstream tasks on defined datasets to analyze ID profiles across layers. Hence, the conclusions drawn in this work are verified across various models, datasets, and tasks, making the findings more convincing.\n\n(2)\tThe comparisons to related works, esp. (Valeriani et al., 2023) which inspires this work, are clearly presented, hence the contributions of this work are clear and solid.\nThe verification of the emergence of a central high-dimensionality phase, and analysis of language processing behavior and performance during the high-dimensionality phase are quite thorough.\n\n(3)\tThe analysis in Conclusion demonstrates that many findings in this work align with prior works and concurrent works. The paper clearly summarizes insights of guidance for future research. The Appendix provides detailed experimental setup and additional results. And finally, the analysis of potential applications of the findings is valuable to the research community.\n\n(4)\tOverall, the paper is clearly written and easy to follow." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work takes a high-level geometric approach to analyze intrinsic dimension (ID) of the representational manifold at each layer of a decoder-only Transformer LLM to understand how layer geometry relates to layer function. Although inspired by the earlier work of (Valeriani et al., 2023), this work greatly extends the models investigated to include five mainstay decoder-only LLMs, and added more extensive probing and downstream tasks on defined datasets to analyze ID profiles across layers. The resulting observations are different from those from (Valeriani et al., 2023). This work made quite a few interesting findings on detecting broad qualitative patterns, and provides useful guidance for future research towards interpretability, analysis of model behavior and quality, and model pruning and layer-specific fine-tuning etc." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "(1)\tAlthough the paper is overall clearly written, please make sure that every symbol used needs to be clearly defined when it first appears, e.g., d in Section 3.4.\n\n(2)\tPlease provide rationale for critical algorithmic designs, for example, please clarify why GRIDE is selected, and why the three alternative measures for comparing layer’s representation spaces are chosen. \n\n(3)\tCurrently, k is still selected based on visual inspection. It would be useful to propose methods that can automatically select k.\n\n(4)\tIt is interesting that OLMo seems a bit of an outlier compared to the other 4 LMs, although it also exhibits the ID peak and other related properties. It would be useful to provide insights on why OLMo behaves differently from the other models, and shed light on patterns of any potential “outlier” LM." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. There seems to have second ID peak in the later layers over LLMs. Do you think this second ID peak might reveal additional insights?\n2. In your analysis (Figure 4), you observed that Pythia and OPT exhibit very similar representations. Could this similarity be attributed to pre-training on similar datasets? If so, how might this influence your findings, and have you considered controlling for dataset overlap to isolate structural factors more effectively?\n3. The work focuses on classification tasks to analyze representation spaces in language models. Could you explain why generative tasks were not included? Do you expect the observed ID peaks and representation patterns to differ in generative contexts?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "This work conducts experiments on various LMs (e.g., OPT-6.7B, Llama-3-8B, Pythia-6.9B, OLMo-7B, and Mistral-7B) using multiple datasets, providing a comprehensive analysis. It also observes how representational intrinsic dimensionality (ID) varies across layers and proposes insightful hypotheses. Furthermore, this work inspires the research community to explore the utilization of ID information in transformer-based LM applications." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper explored how transformer-based language models evolve their internal representations across layers, revealing a distinct high-dimensional abstraction phase. The authors observe the some findings across multiple LMs and datasets, and they provide a foundation for better understanding and optimizing language model architectures. This work bridges the gap between geometric representation and linguistic function in transformer-based LMs. Also, it highlights the potential of intrinsic dimensionality as a tool for analyzing and evaluating LMs." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "While the paper combines two methods, GRIDE (Denti et al.) and Information Imbalance (Glielmo et al., 2020), to analyze four large language models (LLMs), this combination may fall short in terms of novelty. In Section 4.1, the choice of pre-training datasets for evaluation is also a limitation. Since these datasets have likely been encountered by the models during training, the results may not provide a fully accurate picture of the models’ generalization capabilities. Testing on unseen datasets would be crucial to evaluate the robustness and generalizability of the observed patterns, especially in real-world applications where unseen data is the norm. The study is limited to a narrow range of LLMs in terms of scale. Evaluating models of varying sizes (e.g., smaller models alongside large ones) would offer a more comprehensive understanding of how model size impacts intrinsic dimensionality and representation overlap across layers." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "I strongly advice to improve the submission w.r.t. the mentioned weaknesses. That helps both quality and reach.\n\nThe first paragraph of Asset Section C.1 (lines 809-836, in particular 828-829) mentions sensitivity of ID estimation w.r.t. to noise, small scales, density variations and curvature. That analysis suggests some sort of frequency decomposition integrated with the ID estimation." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The findings in bold letters at lines 406-407 and lines 425-426 may be useful to some researchers that need to train and/or select models. The fact that ID seems to change gradually over layers is interesting, but may have a simple explanation in the extreme averaging scale of these models." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The submission is analytic work trying to correlate intrinsic dimension of intermediate NN/LLM representations with linguistic targets. Comparison of the method applied to different models allows insights into some of their learned structural differences." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Except for the few strengths mentioned above, the submission does not explain for what else the gained insights can be used for or wether they are more useful than that at all.\n\nThe analysis focuses only on fully trained models and does not provide insights into how ID changes over time. A correlation analysis to other work would add more value. My first thought was a correlation to the IB method (e.g. Tishby et al. 2000 and Schwartz-Ziv & Tishby 2017), but this may not be the only or best choice.\n\nThe submission wrongly mentions PCA being linear (line 061, applies only to its original form) which leads to the quick conclusion to discard it. This is puzzling as the research on non-linear PCA is quite diverse based on very different techniques and there's even early work using neural networks dating back to 1991 (Mark Kramer: \"Nonlinear PCA Using Autoassociative NNs\")." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "In the data section, it is not very clear to me what does it mean to \"extract 5 non-overlapping partitions of 10k 20-token sequences\" and how the shuffled version is generated, can authors explain more about this?\n\nIn section: The ID peak marks a transition in layer function, I think the relation between ID peak and \\delta(l_i \\to l_first) is not very clear. It has very similar shape in OPT and somewhat in pythia, but LLAMA has a completely different curve. It is maximizing towards the end of the layer instead of the center of layer.\n\nIn section 4.2, authors also claim the relation between ID-peak and a few tasks. However, Figure (a) and (b) do not have very clear co-related trend between ID peaks and tasks' performance. In particular, task performance in Figure 5(b) seems to be monotonically increasing instead of peaking in the middle. Can authors justify more about this?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "I like that authors combine evidence from a few different perspectives to demonstrate the relation between intrinsic dimension peak and transition to abstract processing. They also conduct experiments on a few corpus and a few models as well, which make the claim more general and robust" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work analyzes properties of representations of several LLMs through a few approaches: downstream probing, intrinsic dimensionality and information imbalance. The analysis is mainly developed around intrinsic dimension and they show LLMs typically have a few intrinsic dimension peaks across layers. Additionally, they suggest that those peaks indicate transition to abstract linguistic processing through a variety of analysis" }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The method section is weak and the explanation of intrinsic dimension computing is not enough given its importance in this work. I was not able to identify which variable is corresponding to the intrinsic dimension without going through the cited paper. It seems to be the variable $d$ which authors did not explain what it is.\n\nAdditionally, author made a wrong claim in line 177~178 that \\mu has a generalised pareto distribution. I cannot find any resources claiming this specific distribution is a (generalized) pareto distribution including the original cited paper." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024emergence,\ntitle={Emergence of a High-Dimensional Abstraction Phase in Language Transformers},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=0fD3iIBhlV},\nnote={under review}\n}" }, "abstract": { "value": "A language model (LM) is a mapping from a linguistic context to an output token. However, much remains to be known about this mapping, including how its geometric properties relate to its function. We take a high-level geometric approach to its analysis, observing, across five pre-trained transformer-based LMs and three input datasets, a distinct phase characterized by high intrinsic dimensionality. During this phase, representations (1) correspond to the first full linguistic abstraction of the input; (2) are the first to viably transfer to downstream tasks; (3) predict each other across different LMs. Moreover, we find that an earlier onset of the phase strongly predicts better language modelling performance. In short, our results suggest that a central high-dimensionality phase underlies core linguistic processing in many common LM architectures." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "interpretability", "intrinsic dimension", "large language models" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/44bbf2fa310ab7e9b22953196bd6ec5075ee7736.pdf" }, "presentation": null, "primary_area": { "value": "interpretability and explainable AI" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/ed12e7e0818a442228e9843cb02e29791857cba6.zip" }, "title": { "value": "Emergence of a High-Dimensional Abstraction Phase in Language Transformers" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Thanks for submitting the excellent paper to ICLR. While in general I enjoyed reading the paper, I have a few thoughts on the extension of the paper. Specifically, this paper proposes a new CUDA abstraction that allows users to write new kernels. However, it seems that it is built on top of the fact that all the dimensions should be a multiple of 16. This could be problematic in the context of dynamic shapes where the dimension does not divide 16. Could you please elaborate on how could the proposed technique be extended to such cases?\n\nBesides, the paper uses auto-tuning for further adjust the hyperparameters for a better performance. Could you elaborate how much the tunning overhead is?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "* The paper proposes methods at different levels that simplify CUDA kernel implementations\n* The paper can achieve a similar performance compared to the state-of-the-art implementation" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a new programming library for implementing efficient CUDA kernels. tThe paper contains the three ideas at three different levels for CUDA kernel implementation: (1) At warp-level, the author proposes to organize the tile as the multiple of 16; (2) at thread-block level, the author devises template libraries to overlap between different asynchronous warps; (3) at grid-level, the author proposes methods for managing kernel launch kernel launching overheads. As a result, the proposed library can achieve a performance on par with the existing state-of-the-art implementations." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* The paper has not discussed the tunning overhead with the proposed techniques." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1)\tIs your framework limited to the Hopper series? Can it be applied to A100s, or other GPUs such as the A40/L40?\n2)\tYou focus on the 16x16 register block level, but how can your framework be extended to smaller blocks, such as with GEMV, sparse operations, and masked operations (e.g. non-power-of-two dimensions and strided masking, such as in Natten).\n3)\tThroughout the paper, you focus on BF16 precision (with the exception of softmax); have you considered other data types, such as integer types or floating-point formats like FP8?\n4)\tHow could your framework be extended to handle multi-GPU operations, such as Fully Sharded Data Parallel (FSDP) for split operations? This seems like a natural extension of the producer-consumer model.\n5)\tYou compare yourself against Triton, which also supports AMD GPUs. Can you address this as a potential tradeoff in the paper? Alternatively, if your framework can be trivially extended to ROCm, this should be included in the paper with a demonstration, otherwise it represents a tradeoff between efficiency and portability.\n6)\tYour cost model in Section 2.2 is effectively a Roofline model; could you contextualize this in the existing literature? The results in Table 3 are expected, as reordering increases the arithmetic intensity (FLOPs/Byte) of the inner loops.\n7)\tThroughout the paper, the emphasis on industry versus academic adoption (including the use by undergraduates) feels extraneous and detracts from the main narrative. The paper’s contributions should stand on their own without reliance on external endorsements or applications.\n8)\tFigures 2 and 5 present a simplified sketch for softmax, whereas the true implementation is significantly more complex, potentially leading to a misleading comparison with PyTorch. Furthermore, Figure 2 led me to question why you are using C at all for the API, when the listing could easily have been captured by a python trace (e.g. Triton). This design choice is only clarified upon reviewing the implementation details provided in the appendix and supplementary material.\n\nTo build on these questions, the feedback below addresses specific technical details and aims to enhance overall clarity. While this paper presents a strong contribution toward improving kernel efficiency, addressing these points will better showcase the authors’ contributions.\n\nMinor Technical Errors:\n\n-\t044: The H100 datasheet shows a 7.4x ratio between TCs and ALUs, not 16x. Additionally, my understanding is that the TCs necessarily require bubbles as the Register path cannot keep up with the TC I/O for full throughput. \n-\t136: This should be \"can load\" or \"may load\" instead of \"loads.\" In general, a kernel does not necessarily need to load data from memory. Kernels can rely solely on arguments (loaded into registers at startup) to generate new data. For example, a kernel might generate a pseudo-random noise tensor without accessing memory.\n-\t148: The 32 threads must be within the same quadrant, where “consecutive” or “adjacent” would be more appropriate than “nearby”.\n-\t150: In Ampere, a warp cannot simultaneously occupy different functional units, though separate warps can. For accuracy, please verify this claim against the Hopper documentation or micro-benchmarking paper, otherwise consider omitting if verification is unavailable.\n-\t167: Excess registers spill over into Global Memory, not L1. They can appear in L1 due to the memory hierarchy, but this is at the discretion of the hardware cache manager.\n-\t171: Multiple thread blocks can only schedule on the same SM if there is sufficient space (e.g. SMem), otherwise they would clobber each other.\n-\t173: This statement should be more precise to mention “all thread blocks” and that the L2 is hardware managed, making it distinct from the software managed SMem.\n-\t179: The tail-effect cost mentioned only applies to singular kernels. Ideally the GPU should have multiple kernels in flight, which can run concurrently.\n-\tIt would also be relevant to mention that kernels which contain too many instructions can cause slowdown as they will incur ICache misses.\n\nPresentation Issues:\n\n-\tThe abstract should be revised for clarity, with suggested improvements like “creates a”, “suggest that”, and “resembling PyTorch.”\n-\tThe paper could benefit from clarity revisions in several sections, where phrasing and word choice could make technical details easier to follow. Lines: 073, 170, 178, 205, 278, 299, 301, 328, 370, 397\n-\t325: You should not use \"[1]\" and \"[2]\" to enumerate concepts as they are easily confused with reference indicators. \n-\tTable 2 and Table 3 should probably be Figures like Figure 6. It is also unclear why these stop at 4 stages, K=1024, and what K is. (MxN)x(NxK)? \n-\tFigure 7 and 8 should use subfig captions rather than plot titles. If parameters are common among subfigures, then they should be stated in the figure caption, otherwise in the subfig caption. The fontsize for the axis and labels is too small. Finally, the batch size does not match with the titles and caption.\n-\tThe table in Section 4.2 is missing a caption and column (TK is listed twice).\n-\tThe reference links are broken in Appendix B." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The authors demonstrate significant improvements to computational efficiency within a clearly defined framework that appears relatively straightforward to adapt. Their framework also provides functionality for more complex resource management, which is often challenging to manage directly in CUDA. Additionally, the authors demonstrate the impact of varying hyperparameters for several key kernel operations, most of which match or exceed standard baselines. Lastly, the results show a surprising contrast with Triton implementations, positioning their approach within the CUDA domain while achieving a similar level of complexity to Triton." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a framework to facilitate easy writing of efficient CUDA kernels. The authors leverage the asynchronous compute capabilities of the Hopper series GPUs by following a producer-consumer paradigm, to efficiently overlap different kernel operations. Additionally, the authors investigate the impact of various memory ordering strategies, demonstrating that relatively simple strided patterns offer the best tradeoffs. Lastly, the authors demonstrate performance that is comparable to or exceeds existing methods, including Triton.\n\nOverall, the work provides a significant contribution to improving computational efficiency for common operations, though the application appears limited in scope. Additionally, minor technical and structural errors impact readability. These issues could be addressed in a revision, at which point I would be inclined to raise my score." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "-\tThe application appears limited in scope, which should be explicitly addressed. For example, is the framework limited to Hopper GPUs and above? And the focus on 16x16 register blocks may limit extensibility to other common cases such as GEMV and sparse computations.\n-\tThe paper contains many issues with presentation, including caption errors, grammatical and awkward wording, and typos, all of which impair readability. \n-\tThe paper overlooks relevant computer architecture literature regarding performance modeling, specifically in the context of balancing compute and memory (e.g. roofline analysis). Many of the findings presented in the paper are expected from the existing literature." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "1. What are the fundamental challenges preventing CUTLASS from avoiding bank conflicts? Could it be that the FlashAttention3 kernel simply did not select the optimal layout?\n2. CUTLASS has implemented both ping-pong and cooperative kernel variants for GEMM, with varying performance across different scenarios. How does TK support ping-pong and cooperative kernels, and could you include a comparison with CUTLASS in Figure 7’s GEMM kernel results?\n3. TK appears designed specifically for the Hopper architecture with asynchronous features. Is it also compatible with Ampere or other GPU generations? How does TK’s performance on an A100 compare to Triton?\n4. Following Q3, if Blackwell GPUs were released, would TK’s abstractions remain applicable? How do you plan to ensure extensibility across GPU generations?\n5. What's the usage of the cost model in Section 2.2? This formula is highly simplified and does not guide any optimization or automatic search later.\n6. Section 3.1 discusses various layouts — do users need to manually manage data organization and specify layouts in TK?\n7. Figure 5 is just some wrappers of mbarriers. Any insights here?\n8. Can TK effectively handle quantized kernels, where data layout is crucial for efficient transfers from TMA and WGMMA computation? How does it perform on FP8 GEMM and FlashAttention kernels?\n9. What is TK's performance on causal attention kernels?\n10. Please provide detailed experimental configurations in the Appendix. For example, which versions of PyTorch and Triton were used? Was `torch.compile` employed to optimize those network layers? For cuBLAS, was the latest [cuBLASLt](https://developer.nvidia.com/blog/introducing-grouped-gemm-apis-in-cublas-and-more-performance-updates/) autotuning enabled? Since PyTorch also uses Triton as a backend, what distinguishes the two baselines in Figure 8?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The TK library provides a useful abstraction for writing high-performance asynchronous kernels on GPU.\n2. The presentation is clear and accessible, especially the introductory sections on GPU architecture, which provide a helpful overview for ML researchers who may lack in-depth experience with GPU programming.\n3. The experimental results are compelling, showing performance on par or better than highly optimized kernels, such as FlashAttention3. The paper also demonstrates significant speedups across different kernel types compared to state-of-the-art frameworks like Triton and PyTorch." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents ThunderKittens (TK), a C++ embedded library for writing high-performance CUDA kernels for NVIDIA GPUs. It introduces warp-, thread-block-, and grid-level abstractions to facilitate mapping of kernels to the GPU hierarchy. Experimental results indicate that TK can outperform strong industrial baselines, achieving superior performance for GEMM and attention kernels." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The TK library is still too low-level with too many details, which requires users to manage synchronization carefully and does not simplify the programming burden.\n2. The novelty and advantages of TK over CUTLASS are unclear. Many functionalities seem achievable with CUTLASS as well. The authors mention that TK addresses bank conflicts, but the evidence presented is minimal. There appear to be no inherent limitations in CUTLASS that would prevent it from avoiding bank conflicts.\n3. Similarly, the benefits of TK over Triton are not well established. Triton, embedded in Python with a PyTorch-like API, may offer a more accessible interface. By contrast, TK, embedded in C++, still requires explicit handling of communication with mbarrier operations like expect and arrive. No user study or lines of code comparisons are provided to demonstrate that TK improves programmer productivity.\n4. Experimental results are good, but still missing comparisons in some important cases like quantized kernels and causal attention.\n5. The work reads more like a system paper, with limited ML-focused insights, raising questions about its fit for ICLR.\n\nMinor:\n- P4: \"Since the frameworks are not C++ embedded, it can be challenging to use specialized hardware instructions\" This statement is inaccurate; TVM provides mechanisms to incorporate low-level TensorCore instructions, and Triton also has [inline](https://triton-lang.org/main/python-api/triton.language.html#inline-assembly) operation to include PTX code.\n- Section 2 does not discuss the Tensor Memory Accelerator (TMA) on Hopper, which is essential for asynchronous optimizations mentioned in Contribution 2.\n- Appendix B labels appear broken (??)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Could the authors elaborate on the potential for cross-platform compatibility? Given the focus on NVIDIA’s H100 GPUs, it would be helpful to understand whether TK’s abstractions could be adapted to other GPU architectures, like AMD or Apple, and what challenges might arise.\n\nThe paper demonstrates TK’s strong performance on medium-sized data blocks, but could the authors provide more insights into how well TK scales with very large datasets? Are there specific limitations to consider for applications requiring high parallelization or extensive data handling?\n\nCould the authors expand on their design choice to limit TK to a few key abstractions? Are there specific reasons why additional templates or adaptive features were not incorporated, and would doing so have risked undermining the framework’s simplicity?\n\nIn scenarios with high memory demands, how does TK manage the balance between memory overhead and computational efficiency? Further detail on this balance could clarify TK’s suitability for applications with varied memory and compute requirements.\n\nLastly, could the authors clarify TK’s debugging process, especially for users who may not be familiar with GPU optimization? Since GPU kernel errors can be challenging to diagnose, any insights into how TK might support error handling and debugging would be valuable for potential adopters." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper offers a fresh and practical approach to GPU kernel programming, using only a handful of essential abstractions to make high-performance kernel writing accessible to a wider range of developers. This simplicity-oriented approach can reduce the complexity typically associated with GPU development, which could be particularly valuable for those without extensive CUDA experience. In terms of performance, THUNDERKITTENS shows impressive results, even surpassing established libraries like CuBLAS and FlashAttention-3 in several tasks, especially in backward pass operations for attention mechanisms and linear attention. The results strongly suggest that TK’s design strikes a good balance between simplicity and performance optimization. Furthermore, by aligning its design with PyTorch and NumPy, TK makes it easier for non-specialists to adopt, potentially expanding the accessibility of efficient GPU programming." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces THUNDERKITTENS (TK), a framework that simplifies writing AI kernels for GPUs while still allowing for high performance. Using a few key abstractions, TK provides tools for developers to create efficient kernels without deep expertise in GPU programming. Through benchmarking, the authors show that TK performs on par with or better than other leading frameworks like CuBLAS and FlashAttention-3 for various AI tasks. TK’s accessible design, inspired by PyTorch and NumPy, aims to make high-performance kernel development more straightforward and accessible to a wider audience." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1- While the minimalistic design is a key strength, it may also limit TK’s flexibility for more specialized AI tasks that require tailored optimization strategies. As demands grow for handling complex and emerging AI workloads, the current set of abstractions could potentially fall short.\n\n2- The focus on NVIDIA’s H100 GPUs raises questions about how well TK can transfer to other platforms, such as AMD or Apple GPUs. Expanding on cross-platform compatibility would provide more clarity about TK’s broader usability.\n\n3- Though the paper demonstrates strong performance on medium-sized data, it is less clear how TK handles scalability with very large datasets or highly parallelized scenarios. Addressing its limitations in these settings could further support TK’s value in real-world applications." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "ThunderKittens (TK) is a framework that simplifies the creation of high-performance AI kernels through key abstractions, enabling efficient implementation of ML architectures on GPU hardware and surpassing previous approaches in hardware utilization." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024thunderkittens,\ntitle={ThunderKittens: Simple, Fast, and \\${\\textbackslash}textit\\{Adorable\\}\\$ Kernels},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=0fJfVOSUra},\nnote={under review}\n}" }, "abstract": { "value": "The challenge of mapping AI architectures to GPU hardware is creating a critical bottleneck in AI progress. Despite substantial efforts, hand-written custom kernels fail to meet their theoretical performance thresholds, even on well-established operations like linear attention.\nThe diverse hardware capabilities of GPUs might suggest that we need a wide variety of techniques to achieve high performance. However, our work explores whether a small number of key abstractions can drastically simplify the process. We present ThunderKittens (TK), a framework for writing performant AI kernels while remaining easy to use and maintain. Our abstractions map to the three levels of the GPU hierarchy: (1) at the warp-level, we provide 16x16 matrix tiles as basic data structures and PyTorch-like parallel compute operations over tiles, (2) at the thread-block level, we provide a template for overlapping asynchronous operations across parallel warps, and (3) at the grid-level, TK can help hide the block launch and tear-down, and memory costs. We show the value of TK by providing kernels that match or outperform prior kernels for a range of AI operations. We match CuBLAS and FlashAttention-3 on GEMM and attention inference, and outperforms the strongest baselines by $10-40\\%$ on attention backwards, $9\\times$ on state space models, and $14\\times$ on linear attention." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Systems", "Kernels", "Efficiency", "Efficient Models", "IO Awareness", "GPUs" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/c01e94f1108ac427db7e8cd8e7d916e3feede76a.pdf" }, "presentation": null, "primary_area": { "value": "infrastructure, software libraries, hardware, systems, etc." }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/2b822264ca600550882aee85b36c3fdd96aa88e0.zip" }, "title": { "value": "ThunderKittens: Simple, Fast, and $\\textit{Adorable}$ Kernels" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": null, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": null, "primary_area": null, "questions": null, "rating": null, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": null, "summary": null, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": null, "withdrawal_confirmation": { "value": "I have read and agree with the venue's withdrawal policy on behalf of myself and my co-authors." } }, { "TLDR": null, "_bibtex": { "value": "@misc{\nhuang2024secondorder,\ntitle={Second-order finite-time and fixed-time systems for sparse recovery and dynamic sparse recovery},\nauthor={Yihua Huang},\nyear={2024},\nurl={https://openreview.net/forum?id=0fXcrl35V0}\n}" }, "abstract": { "value": "In the rapidly advancing field of healthcare, efficient processing of sparse data is essential for applications such as medical imaging and personalized medicine.\nThis paper introduces innovative second-order finite-time and fixed-time systems\ntailored for sparse recovery in healthcare data, incorporating control laws into\nthe second-order derivative. We validate the stability and convergence of these\nsystems within finite and fixed times using the Lyapunov method. Furthermore,\nwe examine the tracking performance and assess both practical finite-time and\nfixed-time convergence. The effectiveness of our systems is highlighted through\ncomparative analyses with existing methods, with numerical experiments demon\u0002strating superior accuracy and dynamic tracking capabilities of sparse biomedical\nsignals." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": { "value": [ "~Yihua_Huang3" ] }, "authors": { "value": [ "Yihua Huang" ] }, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "frame of defining penalty functions", "noise", "accelerated distributed generalized reweighted noise filtering consensus algorithm", "accelerated distributed robust generalized reweighted denoise consensus algorithm", "$l_p$-norm minimization", "multi-target tracking" ] }, "large_language_models": { "value": [ "No, not at all." ] }, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": { "value": "huang|secondorder_finitetime_and_fixedtime_systems_for_sparse_recovery_and_dynamic_sparse_recovery" }, "pdf": { "value": "/pdf/2dd49ca54fc6effbed7812f7d412000db391464a.pdf" }, "presentation": null, "primary_area": { "value": "optimization" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": { "value": "No" }, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": { "value": "Yes" }, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Second-order finite-time and fixed-time systems for sparse recovery and dynamic sparse recovery" }, "venue": { "value": "ICLR 2025 Conference Withdrawn Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Withdrawn_Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "- How is the approximated gradient influenced by the depth of the model? I would expect the error to increase as the model gets deeper.\n\nI find the paper very interesting and am ready to increase my grade should my remarks be addressed by the authors." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The paper is well written, clear, and has helpful illustrations.\n- The algorithm seems simple, natural and intuitive.\n- While the algorithm relies on reversible layers, it can still be mixed with standard non-reversible layers, for which a standard backpropagation is performed.\n- The authors validate their algorithm with thorough experiments and analyses." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors propose a new algorithm for training reversible models. Compared to backpropagation, it can be run on each layer in parallel and with a reduced memory cost. They show empirically the advantages of their algorithm on RevNet models for image classification." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Invertible networks are currently not very used. This limits the direct applications of the algorithm. However I am aware that PETRA could motivate such the use of such architectures.\n2. The experiments are only performed on RevNet models for image classification. As mentioned in the conclusion, it would be very nice to see experiments on more tasks and models. Indeed, as PETRA is applicable to only a subset of models (reversible models), it is frustrating to only see experiments on a single architecture.\n3. Lines 509-510: I think you meant RevNet instead of ResNet." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "How PETRA perform on large model and more complex task, such as pretraining language model? The experiment in the paper is weak. The scalability of PETRA can not be verified by the current empirical results. Experiments on distributed pretraining for llm is necessary to validate the efficiency of PETRA, for example: experiments on Pile dataset with varying model size.\n\nIs the reversible architecture necessary for PETRA? For models that integrate both reversible and non-reversible layers, how does PETRA manage memory savings and efficiency, and could these hybrid architectures affect its scalability benefits?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The PETRA paper presents a new alternative for large-scale neural network training, offering efficient parallelization by decoupling forward and backward passes, which enables stages to compute independently across devices. Utilizing reversible architectures, PETRA removes the need for activation and parameter storage, achieving up to 54.3% memory savings, making it especially valuable for training large models. It demonstrates accuracy comparable with backpropagation on datasets like CIFAR-10 and ImageNet." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "PETRA is a model-parallel training method for reversible neural networks that decouples forward and backward passes, eliminating the need for activation or parameter buffers. This enables efficient parallel computation across devices with reduced memory overhead. PETRA matches backpropagation in accuracy on datasets like CIFAR-10 and ImageNet, while also achieving notable speed and memory savings, making it a potential alternative for large-model training." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Dependency on Reversible Architectures: The approach is designed specifically for reversible architectures, which may limit its application to models that can be easily adapted to this structure. Non-reversible architectures, such as standard ResNets or some types of transformers, may not benefit as fully from PETRA’s memory and efficiency gains.\nIncreased Communication Overhead: While PETRA reduces memory usage, its reversible stages require additional communication overhead during the backward pass, which could affect scalability on very large, distributed systems. And the PETRA propose dividing a model into some\nScalability Constraints with Non-Reversible Layers: Although PETRA performs well on reversible architectures, any non-reversible stages still require stored activations, potentially increasing memory use and complicating scalability for models that include such layers." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 4 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "N/A" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The paper demonstrated that activation reconstruction can work well with out-of-sync backward weights, and the reconstructed activations can be used to update weights.\n- The paper has shown real computation and memory savings." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes to perform model parallel training using reversible architectures. Compared to delayed gradient, the proposed method is more memory efficient since it does not need to stash weights. It is shown that on shallower architecture the performance is slightly better than regular backprop and on deeper architecture such as ResNet-50, there is a slight drop but not significant. Overall, the work is likely to have a big impact as a way to scale up model parallel training." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- It would be nice to see at what scale the method starts to break down (say when there is more and more delay in reconstruction). And show a plot on reconstruction error and final performance as a function of the number of delay steps. The model depth can be another variable to explore, aside from the few standard model architectures, perhaps sweeping a wider range of depths.\n- Algorithm 1 is a little hard to process.\n- The method relies on gradient accumulation to fully match with the It is unclear to me how gradient accumulation would have any impact when a large batch / data parallel is employed. This may not be a concern for LLMs, but for ImageNet and SSL training, many use very large batch sizes." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "How did you partition the architectures for your experiments? How many layers/blocks in each stage? Were they all the same size? And if so, would that not bring them out of sync during training such that top layers/stage were idle a lot of the time? The size of the feature maps is decreasing in the layer index, no? Thus, the lower layers/stages would consume more memory and compute than the top ones?\n\nPerhaps you could add some information about this in the appendix :-)" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper is well-written and easy to follow. The idea of utilizing reversibility for parallelization is a nice, simple, and novel idea! Consequently, I find myself sufficiently convinced that the method works --- albeit, that the empirical evaluation is somewhat limited. The novelty and applicability of the method mostly outweighs my concerns about the evaluation." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes to utilize the concept of reversible architectures to improve parallelization in DNN training. A model is split into multiple stages that are trained asynchronously; i.e. in a model parallel fashion. Leveraging reversibility, the training of the different stages is effectively decoupled. This scheme offers a linear speedup in the number of stages relative to end-to-end backprop, while reducing the memory footprint. The method is evaluated using ResNets/RevNets with three different image classification benchmarks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "My only objection to this work is the limited number of experiments. They are limited to ResNet/Revnet 18/34/50 and CIFAR10, ImageNet-32, and ImageNet. It would definitely improve the paper to have at least a few more architectures included." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We show how combining reversible architectures with delayed gradient approaches for model parallelism allows to achieve computational speedups with drastic memory reduction." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024petra,\ntitle={{PETRA}: Parallel End-to-end Training with Reversible Architectures},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=0fhzSFsGUT},\nnote={under review}\n}" }, "abstract": { "value": "Reversible architectures have been shown to be capable of performing on par with their non-reversible architectures, being applied in deep learning for memory savings and generative modeling. In this work, we show how reversible architectures can solve challenges in parallelizing deep model training. We introduce PETRA, a novel alternative to backpropagation for parallelizing gradient computations. PETRA facilitates effective model parallelism by enabling stages (i.e., a set of layers) to compute independently on different devices, while only needing to communicate activations and gradients between each other. By decoupling the forward and backward passes and keeping a single updated version of the parameters, the need for weight stashing is also removed. We develop a custom autograd-like training framework for PETRA, and we demonstrate its effectiveness on standard computer vision benchmarks, achieving competitive accuracies comparable to backpropagation using ResNet-18, ResNet-34, and ResNet-50 models." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Model parallelism", "Delayed gradient", "Reversible architectures" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/3d00e6616b7a6732f6741738818efb89f355c8e8.pdf" }, "presentation": null, "primary_area": { "value": "optimization" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/f014b7dfedb72aeef80dcbbd894dd8047ec65620.zip" }, "title": { "value": "PETRA: Parallel End-to-end Training with Reversible Architectures" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "I would appreciate the authors response and clarification on the points listed under \"weaknesses\"" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "Strengths:\n- The topic is interesting and of recent interest\n- The approach (CAP and PPP) seems novel.\n- The experiments show significant gains over the baselines in deriving penalty heuristics for guided local search, as well as more moderate gains on constructive heuristics for TSP, heuristic measures for ant colony optimization, and reshaping of attention scores in neural combinatorial optimization," }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents Hercules, an LLM-based algorithm for generating heuristics for combinatorial problems. The paper seems to extend the framework in Ye et al., 2024 with a more advanced direction generation (based on identifying core components in heuristics) as well as an LLM-based fitness calculation. The experiments show gains over the baselines." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Weaknesses:\n- I found the claim about information gain to be quite confusing. \n\t- First, a lot of information is missing: why the number of core components corresponds to the number of heuristics (can we not have multiple core components per heuristic or the same core component in multiple heuristics)? why do we assume that the set of all possible directions can be partitioned into mutually exclusive subsets that correspond to components (can we not have the same direction for multiple core components)?\n\t- Second, it is really not clear why the information gain means we get better heuristics (as indicated in lines 284-285)? If the components generated are of low-quality the directions may be of lower quality as well.\n\n- Experimental evaluation:\n\t- It is not clear what is being reported under gain: the definition is based on \"the performance of ...\" but it is not clear how performance is measured.\n\n- Writing: the writing could improve as a lot of information is not clearly presented. For example there are no clear definitions for a range of terms like parent heuristics, elite heuristics, etc.\n\n- The paper does not provide significant insight into the impact of the proposed techniques (CAP and PPP) beyond the experimental results. For example, it would be interesting to show an analysis of the correlation between predicted fitness values and quality of heuristics." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Could you provide details on the outer meta-heuristics (GLS, ACO etc.)? How much of the results are due to the LLM integration with CAP, PPP etc. vs. the meta-heuristics leading the search into good solutions. \nIt would be interesting to know the comparison between default GLS, ACO or even other baselines for TSP, BinPacking, MKP to position the results in this paper. As is, it is hard to evaluate the significance" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "Integrating LLMs with heuristic solving is an exciting combination. \nThe paper implements an end-to-end pipeline that starts with a seed query that then mimics evolutionary computing via LLMs, yielding heuristics that can be embedded in the Local Search Meta-Heuristics for different combinatorial problems. \nThe connection with Information Gain is an excellent addition\nFrom a practical perspective, the paper considers several details into account such as reducing costs via LLM predictors." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper studies the generation of heuristics for combinatorial optimization problems using LLMs. \n\nThe work continues similar work in this space that tries to mimic evolutionary computation (crossover, mutation) via LLMs. The result is an LLM-infused metaheuristic algorithm. \n\nUnlike the previous work, the paper claims 1) to address introducing more problem specificity into the prompts and 2) to speed up the process by using LLMs to predict the performance of generated accuracies to skip over evaluating them fully. \n\nOverall, I enjoyed reading this paper, and I appreciated the work that went into building an end-to-end pipeline with several components." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "As rightly noted in the paper, the idea of mimicking evolutionary computation via LLMs is not new. In fact, most (all?) crossover and mutation operators are from Ye et. al. 2024. On the one hand, the experiments and the ablation study show that the proposed modifications might offer some benefit in the results, and on the other hand, they can be regarded as incremental, and it is not clear what's the main takeaway. \n\nRegarding the presentation, I found it difficult/confusing that many moving parts are introduced as large components with several acronyms Hercules, CAP, PPP, EXEMPLAR, Cons --but after all, the provided pseudocode shows the overall algorithm, so I am not sure what these abstractions add to the presentation. Also, the paper claims our \"propriety\" CAP algorithm --what does that mean? \n\nThe idea of adding more specificity to the prompts seems reasonable at a high level, but the paper overindexes too much into the example in Figure 1. The information gain analysis is interesting (and is borrowed from a Hu et. al. 2024) but at the end what happens is we select top-k core components. And that's also not uniform, we do that only some number of iterations (denoted by \\lambda in the paper), all of which remain as more hyper-parameters to deal with.\n\nThe experiments cover TSP, CVPR, Binpacking, Multi-Knapsacks. Importantly, the starting seed function seems critical to the approach. The method generates heuristics but the overall approach to solve these problems are meta-heuristics. (please correct me if I understand this correctly). For TSP, we use guided local search. For BinPacking and Knapsacks we use Ant-Colony Optimization. One might argue that the settings of the outer meta-heuristics and their performance are crucial to the overall results and not just the heuristics (generated by LLMs here.) The experiments do not discuss or study any of this. \n\nAdditionally, all comparisons are with other LLM-based heuristics generations. Note that this is quite a costly approach (hence some effort with performance predictors to save time etc.). According to the tables in the appendix, we are consuming many many minutes upto 5 hours. Then, it is not clear to me how to fairly evaluate these results. How does the same GLS and ACO without the advanced heuristics found by LLM but with standard heuristics perform given the same amount of time? (Btw, does this time include LLM queries or only running the heuristics after the LLM generates them against the instances?) \n\nThis might not be surprising that the choice of LLM quite affects the results (Table 1; LLama vs GPT-4o). But then it makes one wonder how much of the value comes from the many moving components proposed here vs. plain and simple, the underlying LLM." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "The differences are small in some cases and it would be great if the authors could\nprovide error bounds or confidence intervals for the empirical results.\n\nWhy is KGLS is reasonable base heuristic?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The proposed framework is interesting and seems to work well in practice." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a framework to use LLMs to generate heuristics for solving\noptimization problems. The authors describe their framework and evaluate it\nempirically, comparing to other approaches in the literature." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The choice of KGLS as seed heuristics should be justified as it was not designed\nfor the general TSP. Why not LKH? This should also be considered in the\nempirical evaluation; in particular to answer the question of whether KGLS is a\nreasonable heuristic to start with in this case (improving over a weak heuristic\nis easier than improving over a strong heuristic).\n\nFigure 5 has no axis labels." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "This submission does not have ethics concerns." }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please reply to my comments in \"Weaknesses\"." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The paper is laudable for its well-structured and logical presentation, providing a comprehensive understanding of the research topic.\n\n2. The article is praiseworthy for its extensive experimental data and significant findings. The authors have selected a number of baselines for comparative experiments on different benchmarks.\n\n3. The supplement provided in this article is adequate. It explains in detail for the reader what is not expanded in detail in the paper, including specific experimental data, hyperparameter settings, Critical Difference Analysis, etc." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper explores the application of LLM in autonomously generating heuristics to solve COPs and proposes a novel algorithm named Hercules to address the two main challenges of existing approaches.\n\nHercules utilizes the Core Abstraction Prompting (CAP) method to abstract core components from elite heuristics and incorporate them as prior knowledge in prompts, thereby reducing the specificity of search directions. This paper further introduces Hercules-P, an efficient variant of Hercules that integrates CAP with the novel Performance Prediction Prompting (PPP) method. PPP leverages LLMs to predict the fitness values of newly derived heuristics based on their semantic similarity to previously evaluated ones, significantly reducing the required computing resources." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. In the literature of TSP and CVRP, it is known that those conventional heuristic algorithms, such as LKH [1] and EAX [2], exhibit robust performance. It appears, however, that this submission does not address LKH and EAX, nor does it provide a comparative analysis of the proposed algorithm against these established methods.\n\n1. In lines 85-100 of the Introduction section, the authors describe two challenges to LLM-based HG methods, mentioning in the second challenge that these methods introduce numerous linear operations and conditional branches that make the GPU less efficient for these algorithms. In lines 113-128, the authors claim to have proposed Hercules-P in order to better address the second challenge, but I don't seem to have read in the manuscript how Hercules-P reduces linear operations and conditional branches, making GPUs more efficient in processing these algorithms. May I ask if the authors have solved this challenge? If not, these representations are inappropriate.\n\n2. Does the appearance of CVRP in Section 4.4 stand for Capacitated Vehicle Routing Problem? The authors do not explain what CVRP stands for in the body of the manuscript, and the only explanation appears in the code comments in the Appendix section (line 1337). This cannot be very clear to the reader when it comes to understanding the manuscript.\n\n3. In Section 2.3, the authors mention two challenges for NCO solvers: improving generalisation capabilities and large-scale COPs performance. In Table 5, for LEHD, the performance improvement of either Hercules or Hercules-p gradually decreases as the problem size of TSP or VCRP increases. Does this mean that Hercules also fails to address the challenges faced by NCO solvers? Can Hercules still provide performance gains when the problem size is larger? Further discussion is requested from the authors.\n\n\n## References\n[1] Keld Helsgaun. General k-opt submoves for the Lin-Kernighan TSP heuristic. Mathematical Programming Computation 1(2-3): 119-163 (2009)\n\n[2] Yuichi Nagata, Shigenobu Kobayashi. A Powerful Genetic Algorithm Using Edge Assembly Crossover for the Traveling Salesman Problem. INFORMS Journal on Computing 25(2): 346-363 (2013)" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024efficient,\ntitle={Efficient Heuristics Generation for Solving Combinatorial Optimization Problems Using Large Language Models},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=0fwJMANq9P},\nnote={under review}\n}" }, "abstract": { "value": "Recent studies exploited Large Language Models (LLMs) to autonomously generate heuristics for solving Combinatorial Optimization Problems (COPs), by prompting LLMs to first provide search directions and then derive heuristics accordingly. However, the absence of task-specific knowledge in prompts often leads LLMs to provide unspecific search directions, obstructing the derivation of well-performing heuristics. Moreover, evaluating the derived heuristics remains resource-intensive, especially for those semantically equivalent ones, often requiring unnecessary resource expenditure. To enable LLMs to provide specific search directions, we propose the Hercules algorithm, which leverages our designed Core Abstraction Prompting (CAP) method to abstract the core components from elite heuristics and incorporate them as prior knowledge in prompts. We theoretically prove the effectiveness of CAP in reducing unspecificity and provide empirical results in this work. To reduce the required computing resources for evaluating the derived heuristics, we propose few-shot Performance Prediction Prompting (PPP), a first-of-its-kind method for the Heuristic Generation (HG) task. PPP leverages LLMs to predict the fitness values of newly derived heuristics by analyzing their semantic similarity to previously evaluated ones. We further develop two tailored mechanisms for PPP to enhance predictive accuracy and determine unreliable predictions, respectively. The use of PPP makes Hercules more resource-efficient and we name this variant Hercules-P. Extensive experiments across various HG tasks, COPs, and LLMs demonstrate that Hercules outperforms the state-of-the-art LLM-based HG algorithms, while Hercules-P excels at minimizing computing resources. In addition, we illustrate the effectiveness of CAP, PPP, and the other proposed mechanisms by conducting relevant ablation studies." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Heuristic Generation", "Large Language Models", "Combinatorial Optimization Problem" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/ac3fa83ee38011cab5c926ee7d3021d8e5560d77.pdf" }, "presentation": null, "primary_area": { "value": "generative models" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Efficient Heuristics Generation for Solving Combinatorial Optimization Problems Using Large Language Models" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 2 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "How does the SWDI method compare to deep learning-based approaches or other adaptive wavelet techniques in terms of accuracy and computational efficiency?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "Originality\nThe proposed SWDI method creatively combines wavelet analysis with differential inclusion to address limitations in current shrinkage methods by focusing on both strong and weak signals, thus enhancing the detection of signal features important for clinical applications.\n\nQuality\nThe paper is well-founded with rigorous theoretical analysis that supports the authors' claims.\n\nClarity\nOverall, the paper is clearly written, although some improvements can be made (see 'Weaknesses' section)." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents a novel method, the Splitted Wavelet Differential Inclusion (SWDI), for enhancing neural signal analysis, particularly for applications related to Parkinson’s disease. SWDI introduces a dual-parameter approach that estimates both the strong and whole signals simultaneously, addressing limitations of previous wavelet shrinkage techniques. The authors demonstrate that their closed-form solution path improves estimation accuracy for both signal components. This work contributes to the field of neural signal processing by offering a robust framework for analyzing complex neural data, which could have significant clinical applications in neurodegenerative disease research." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The clarity of presentation of this paper can be improved. There are some English mistakes, subject-verb disagreement, missing conjunction 'and', etc. These should be carefully addressed prior to the publication of this paper.\n\nExamples:\n\nline 052: Add 'and' before 'non-parametric shrinkage'.\nline 053: Change 'contains in the signal' to 'contained in the signal'.\nline 056: Change 'composed by' to 'composed of'.\nline 073: Change 'On the other' to 'On the other hand'.\nline 107: Add 'and' before 'non-parametric shrinkage'.\nline 114: Change 'have' to 'has'.\nline 123: Add 'and' before 'the non-burst component'.\nline 373: Change 'includes' to 'include'.\nline 377: Change 'the same ... with' to 'the same ... as'.\nline 382: Change 'as a contrast' to 'in contrast'.\nline 386: Change 'compare to' to 'compared to'.\nline 402: Add 'and' before 'then increases'.\nline 500: Insert 'be' between 'may' and 'due to'.\nline 512: Change 'Fig. 3,. 4' to 'Figs. 3 and 4'." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "Medical data from Parkinson patients is used." }, "flag_for_ethics_review": { "value": [ "Yes, Discrimination / bias / fairness concerns" ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Please see the weaknesses and reply to those." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "* It is a nice idea to go beyond considering only large wavelet coefficients, i.e., the strong part of the signal.\n* The author substantiated her/his novel approach with a theoretical foundation (see, i.e., Theorem 4.6)." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper improves on the well-known wavelet shrinkage approach, introducing a novel method coined \"Splitted Wavelet Differential Inclusion (SWDI)\". as opposed to wavelet shrinkage, it also takes weak components of the signal to be analyzed in consideration. The effectiveness of SWDI is showed by numerical experiments on data from Parkinson patients." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* At present, we have powerful methods based on learning. It is not clear and now even discussed why those are not taken into consideration, There might be good reasons, but this requires a careful discussion.\n* The numerical experiments only compare to other model-based methods, mainly wavelet-based approaches. Again, in partiular, learning based methods (DNNs, etc.) need to be used for comparison." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "Clarification is needed regarding the problem setting, statistical assumptions, and the choice of baseline methods (see weaknesses above)." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The considered problem is central and important.\n- The application to neural signals, specifically in the context of Parkinson’s disease, is interesting." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper considers the longstanding problem of recovering a temporal univariate signal from its noisy observations, which a fundamental problem in signal processing. The authors approach this challenge using wavelet analysis, where they propose to partition the signal into strong and weak components based on the magnitudes of the wavelet coefficients. The paper presents a new method, termed Splitted Wavelet Differential Inclusion (SWDI), which is designed to recover the strong component by employing a differential inclusion framework. It is shown theoretically and empirically in a simulation that the proposed method recovers the strong signal more accurately than other methods based on wavelet shrinkage. Additionally, the method is demonstrated in application to neural signals, where the goal is to identify medication effects on Parkinson’s disease." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The **presentation** of the entire paper, and particularly, the technical aspects is challenging to follow, which hinders comprehension of the core ideas and derivations.\n- Due to the presentation style, it is difficult to clearly appreciate the novelty of the paper. The mix of formal and informal statements complicates rigorous validation.\n- The **problem setting** lacks clarity, particularly its statistical model. While the strong signal is defined based on the noise standard deviation $\\sigma$, the method’s dependence on the signal-to-noise ratio (SNR) or $\\sigma$ is unclear. Additionally, it is not specified whether the derivations and results assume Gaussian noise or if the true signal $f$ is deterministic.\n- **Numerical results** are limited. Figures are of low resolution with small fonts, making them hard to interpret, especially Figure 1. Expanding the numerical experiments to encompass a broader range of cases and a more comprehensive comparison with alternative shrinkage methods would enhance the paper. The justification for the selected baselines is unclear, considering the prevalence of other methods addressing the same problem." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 2 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "It would be beneficial for the authors to discuss the tradeoffs between wavelet techniques and more recent methods, such as deep learning-based approaches, in this specific application. A comparison or discussion of how their method performs relative to recent non-wavelet techniques would provide valuable context for evaluating the method's effectiveness.The distinction between weak signals and noise could be clarified further. I suggest that the authors provide more detailed criteria for how they distinguish between the two, and discuss whether there are alternative approaches. The \"differential\" aspect of the proposed method requires clearer explanation. A step-by-step description of how the differential inclusion is applied, ideally with a simple example, would greatly improve the clarity of this concept and make it more accessible to readers." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The paper introduces the Splitted Wavelet Differential Inclusion (SWDI) method, which improves the estimation of both strong and weak neural signals by utilizing an ℓ2 splitting mechanism. It demonstrates better accuracy than traditional wavelet shrinkage, particularly in Parkinson's disease signal analysis, capturing non-burst activity alongside stronger signal components." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes the Splitted Wavelet Differential Inclusion (SWDI) method for neural signal processing, achieving better strong and weak signal estimation in Parkinson's disease analysis." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "To better demonstrate the proposed method's practicality, I suggest including comparisons with non-wavelet methods, particularly those that have seen recent success in this field. This could provide a clearer perspective on how the method performs in a wider range of real-world applications. Specific examples of non-wavelet techniques, such as deep learning-based methods, would strengthen the evaluation. While wavelet techniques have been widely used in the past, it would be helpful if the authors could justify their choice of wavelets in this context and explain how their approach advances the state-of-the-art. Additionally, comparisons with more recent works in the field would help to clarify the method's relevance and novelty. The paper's content seems more suitable for signal processing journals or conferences, such as TSP, INDIN, or ICASSP." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose a new Wavelet smoothing method to enhance the signal reconstruction in neuroscience applications." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024splitted,\ntitle={Splitted Wavelet Differential Inclusion for neural signal processing},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=0gGPVbRqOE},\nnote={under review}\n}" }, "abstract": { "value": "Wavelet shrinkage is a powerful tool in neural signal processing. It has been applied to various types of neural signals, such as non-invasive signals and extracellular recordings. For example, in Parkinson's disease (PD), $\\beta$ burst activities in local field potentials (LFP) signals indicated pathological information, which corresponds to \\emph{strong signal} with higher wavelet coefficients. However, it has been found that there also exists \\emph{weak signal} that should not be ignored. This weak signal refers to the set of small coefficients, which corresponds to the non-burst/tonic activity in PD. While it lacks the interpretability of the strong signal, neglecting it may result in the omission of movement-related information during signal reconstruction. However, most existing methods mainly focused on strong signals, while ignoring weak signals. In this paper, we propose \\emph{Splitted Wavelet Differential Inclusion}, which is provable to achieve better estimation of both the strong signal and the whole signal. Equipped with an $\\ell_2$ splitting mechanism, we derive the solution path of a couple of parameters in a newly proposed differential inclusion, of which the sparse one can remove bias in estimating the strong signal and the dense parameter can additionally capture the weak signal with the $\\ell_2$ shrinkage. The utility of our method is demonstrated by the improved accuracy in a numerical experiment and additional findings of tonic activity in PD." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Wavelet smoothing", "differential inclusion", "weak signal", "signal reconstruction", "Parkinson's disease", "burst activity" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/04306f7c92f670b242d9177a8b5c6bfa7f07858b.pdf" }, "presentation": null, "primary_area": { "value": "applications to neuroscience & cognitive science" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Splitted Wavelet Differential Inclusion for neural signal processing" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. In figure 7, ViTARC-VT has very large variance in terms of performance. Any reason for it? Also what is the key technique leading to the significant performance improvement from ViT-Vanilla to ViTARC-VT? BorderTokens does not look to be important from this figure. Is this because of the 2D positional encodings in ViTARC-VT, instead of 1D positional encodings in ViT-Vanilla?\n2. The Equation (12) is not quite clear. I did not find how to calculate the value of $r_{left}$ and $r_{right}$.\n3. It is unclear how to tune the hyper-parameters $\\alpha$ and $\\beta$ in Equation (10), and no ablation studies are provided." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The Abstract Visual Reasoning (AVR) tasks is interesting and important to study because it requires strong reasoning capability of ViTs. The paper also has very interesting findings, highlighting the importance of positional encodings in solving pure vision-based visual reasoning tasks.\n2. This works provides very detailed model improvements they have tried to improve the performance, from ViT-Vanilla to ViTARC-VT and ViTARC. This is very useful to the communities to reproduce the experiments and improve further.\n3. The final model ViTARC achieves very strong performance in most of the tasks, which is also a significant improvement over ViT-Vanilla." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper studies vision transformer in abstract visual reasoning ARC tasks which do not include any text or background knowledge, and focus purely on visual abstraction and pattern recognition. However, directly training a ViT with one million examples per task fails on most ARC tasks. Then, several techniques are proposed to improve model performances, including improved 2D positional encodings, and object-based positional encoding. This work highlights the importance of positional information for utilizing Vision Transformers in visual reasoning tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The training/evaluation protocol is not clearly defined. The paper does not show clearly the generalization ability on unseen tasks. All the task they use for evaluation have some training examples during training. It would be very interesting to use some tasks pure for evaluation which are not seen during training.\n2. Some of the key techniques used in this work are not new, like 2D (Relative) Positional Encoding, which have been discussed in the original ViT/Swin Transformer papers and plays a key role for performance improvement in this work. Though some new techniques are introduced in this work, like Positional Encoding Mixer (PEmixer) and Object-based Positional Encoding (OPE), the overall contribution and novelty is marginal." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1 - Did you try using RoPE embeddings instead of RPE, in the second point of Section 5? I am curious about the differences.\n\n2 - About RPE: the paper mentions there is an encoding for left and above, and another one for right and below. How are \"above and to the right\" patches encoded? Why not using an explicit \"above left\", \"above right\", \"bottom left\", \"bottom right\"? It seems like the approach is re-using the 1D sequential approach without taking into account that we are modeling 2D inputs.\n\n3 - How do your architectural changes influence other vision tasks? E.g. classification, detection, VQA, etc. The SOTA ViT models are nowadays used for many tasks (at least as part of many tasks as vision encoders). It would be great if the changes proposed in the paper did not hurt performance in other tasks.\n\n4 - Related to previous question, did you try starting from a pre-trained ViT?\n\n5 - Did you try training jointly for all tasks? Adding a task ID as context for example.\n\n6 - Did you notice any overfitting or underfitting on the final models? Any scaling properties with the data? On the final models, is 1M samples necessary, of 100k would be enough? Would the models perform better with even more samples?\n\n7 - When generating pixels, during evaluation does the pixel value have to be exact? Is the prediction done as a classification in the RGB [256 x 256 x 256] space? Or is it a regression in the [0-1](x3) range? If it is a regression, how is the correctness evaluated?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1 - The paper addresses an interesting question, which is: if the tasks in ARC are visual tasks, how can se use the current vision tools to deal with them? \n\n2 - The reasoning behind every contribution in the paper is well explained. The paper is easy to follow. \n\n3 - Related to the previous point, I particularly like the analysis in Figure 6. It helps understand what the model is (not) paying attention to.\n\n4 - The results in the paper show a clear improvement with respect to the original ViT vanilla baseline, meaning the proposed contributions, overall, are helpful." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper addresses the ARC benchmark from a vision perspective, and answers the question of what changes are required in current SOTA vision architectures to deal with the tasks in ARC. \n\nThe paper suggests that the bad results of vanilla ViTs on these tasks are due to their poor encoding of spatial and positional information, and proposes approaches to deal with this limitation.\n\nFinally, the paper shows results that indicate that the proposed changes were useful for the performance on the ARC benchmark." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The paper has some weaknesses that I believe can be addressed, but also should be addressed.\n\n**1 - Unclear that this is vision modeling**.\n\nThe main argument of the paper is that vision transformers should be improved to work on ARC. But the proposed changes make the model not be a vision model anymore. While the paper mentions that \"Transformer-based LLM approaches convert the images into strings, which does not fully capture all relevant structural information”, this is not too different from what this paper does. \n - By construction, vanilla Transformers model inputs as sequences, but there are some aspects that make them more \"vision\" specific. Some of them are not included in the original ViT (e.g. local attention), and the others (e.g. patching) are removed in this paper (see next points). A non-Transformer-based architecture (e.g. a U-Net in a diffusion model) would make the connection to vision more clear. (I'm not suggesting such a model should be used, just exemplifying the point).\n - The pixels are encoded in a 1x1 grid, effectively making it a sequence.\n - There is an end-of-sequence token, just as there would be if the task was modeled as a sequence. \n - Even object positional encodings are added, abstracting away the low-level vision from the tasks. \n\nAn emphasis on vision is given in the paper more than once, e.g. \"However, in vision tasks, especially those requiring detailed visual reasoning, spatial relationships often carry as much importance as, if not more than, the content of the tokens\". I believe, however, that a lot of the learned lessons are not about vision, but about structured predictions. In general vision tasks, for example, the content of tokens *is* more important than the spatial relationships.\n\n\n**2 - Unclear significance of contributions**.\n\nOverall, the paper reads as a sequence of contributions the authors tried, one after another, and building on the previous one, without any global understanding of the problems. I believe the process to get to a solution should not be part of the paper. The paper should just present the final results and justify them. This makes the presentation confusing, for example when showing results in section 4, before going back to explaining some more technical contributions in section 5. \n\nBut the main problem is not about presentation. I believe there is some disconnect between the different contributions. Some of the latter ones should make the former ones unnecessary, and these are not ablated properly. The following are some examples:\n\n - First, learnable embeddings are removed in favor of positional embeddings. But then a PEmixer is necessary to learn position-specific embeddings. And also, RPE is required because APE encodes spatial changes very slowly. All of this is confusing. I believe the paper should directly start by explaining the final encoding and its reasoning, not present two different encoding techniques before the final one and show how they are not ideal.\n - The paper presents quantitative results showing that PEmixer and OPE, on top of ViTARC-VT actually *decrease* the performance of the model. \n - The padding is added at the beginning, but then it results in problems that need to be corrected. I address this in more detail in the next weakness.\n\n**3 - It is unclear that all the padding contributions are necessary**.\n\nWhy are padding tokens necessary? The original formulation (sequential padding at the end), where the padding tokens are ignored by the attention should be the correct one (computationally it also makes more sense). The per-row padding (without being attended to) is effectively the same as the per-sequence padding, so I am not sure why it is being mentioned. \n\nI understand that there is a problem about the output not understanding boundaries corrected. But the per-row EOS tokens should be enough to address this issue. For the model, it should be equally hard to predict an \"end of row\" token than it is to predict the current _<arc endxgrid>_ token. Would it be possible to ablate this? This is another example of weakness #2. The final solution should not include previous iterations of the solution, if these are not necessary. No (attended-to) padding would imply: \n - More efficient forward and backward passes.\n - No need for three different kinds of end-of-grid tokens. Only a single one would be enough.\n\n**4 - No baselines or comparisons**.\n\nThere are no baselines or comparisons to other approaches, only to their own vanilla ViT. Also, I could not find the paper's performance on the private set of the ARC benchmark, so it is hard to compare to SOTA approaches. Specially interesting comparisons would be to approaches that model ARC as a sequence using an LLM, as I believe the approaches are not very different (see weakness #1).\n\n---\n\nOther minor weaknesses:\n\n- Figure 8 Appendix A seems to motivate the first main technical contributions of the paper. It should not be in the appendix.\n\n- It is unclear why the PEMixer is helping. If every example in the task has different coordinates that are important, I don't understand why it would learn a global (across all samples) weighing. What would be the benefit of giving one position more weight than another one, if every example has different important positions?\n\n- The paper could contain more clarifications and ablations. See \"Questions\" section." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. Can you show that the model works on other complex reasoning tasks outside of ARC as mentioned in point 1 above?\n2. Does the model work without strong priors, e.g., on input and output grid shape?\n3. Can you add in more baselines trained on the same amount of data?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "I appreciate that the authors explore the limits of a data-driven approach to ARC, as well as propose potential inductive biases to encode into a reasoning model for ARC. General priors for reasoning tasks are indeed important. The quantitative results compared to a naive ViT are promising for ARC." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors focus on a data-driven model to solve ARC. They first establish that vanilla ViTs fail on ARC despite being trained on a million examples. They then show that a 2D visual representation with ViT improves performance, and that object-based positional information further yields significant performance gains on ARC." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Many of the architectural designs in the proposed model are made for solving ARC specifically. I believe ARC is a great intermediate proxy task for complex reasoning, but should not be an end goal in and of itself. With enough inductive biases, I believe that solving ARC with a million examples is reasonable, but is not particularly enlightening for the community. For example, 2D padding with <arc_pad> tokens and border tokens <arc endxgrid> that define grid conditions, etc, are very much defined for ARC itself. Would like to see if this model generalizes to other reasoning tasks, for example Raven's progressive matrices, odd one out challenges, etc. \n2. In addition, I'm not convinced that these are indeed the best inductive biases. For example, I believe that by using a \"2D template\" indicated by padding, border, and new-line tokens, the method is endowed with a priori knowledge of what the final grid shape should look like (and also, what the initial grid shape looks like). One core challenge of ARC is precisely that it needs to infer what the output dimensions are (how the shape transforms). Giving the model knowledge that one token should represent one block in ARC is a strong prior to be injecting. \n3. The object-based positional encodings based on OpenCV's contour detections will struggle on more complex or different shapes. This “objectness” should be captured by the visual tokens implicitly. \n4. No other baselines than ViT are explored, though there have been many works proposed for ARC & related reasoning tasks." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "What do the authors think about adding a caveat in the paper about W2 above? (i.e. the fact that this makes zero progress on the ARC challenge, and that the benefits of the proposed methods still need to be demonstrated on a meaningful task)\n\nI can't solve Task A (6e02f1e3) in Fig. 2. I guess other readers would benefit from an explanation? (I'm afraid this sortof makes a Vanilla ViT superhuman on this task!)" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The paper studies a popular architecture (ViTs), so it should be of broad interest.\n\n- The paper shows a domain where ViTs fail dramatically. The task itself is not interesting (ARC in a new, supervised, large-data setting), but the finding is interesting because it points at intrinsic deficiencies of ViTs.\n\n- Several modifications are described that improve the performance. It's kindof the opposite of an ablation study, but it does the job of demonstrating that novel methods imbue ViTs with novel capabilities." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper studies the ARC benchmark, assessing the potential of ViTs for this task when trained with supervision on a single task at a time with plenty of (procedurally-generated) data (which is different from the original few-shot setting).\n\nThe paper first shows that standard ViTs fail. It then shows several modifications to the architecture that make them perform much better." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I don't see significant flaws in this paper. Potential weaknesses:\n\n- (W1) The proposed modifications (in particular the visual tokens) are quite straightforward. This can be seen a good thing. I am actually surprised that the absolute PE and padding (to handle images of different sizes or aspect ratios) have not been used before. Are the authors certain this hasn't been described in the existing literature?\n\n- (W2) There is a valid argument that this paper solves a task that is extremely uninteresting in itself. It has no analogue in the real world and it completely defeats the purpose of the original ARC challenge (because of the supervised, large-data setting, focusing on a single task at a time). This paper makes absolutely no progress towards the goal of the ARC challenge. I still think that the findings are interesting for the reasons mentioned in my \"Summary\" above, i.e. that the proposed improvements give ViTs new abilities. The main issue now is that the contributions of this paper will only have value if/when the abilities are demonstrated to be useful for another task/setting." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "Getting vision transformers to \"reason\" on the Abstraction and Reasoning Corpus (ARC)" }, "_bibtex": { "value": "@inproceedings{\nanonymous2024tackling,\ntitle={Tackling the Abstraction and Reasoning Corpus with Vision Transformers: the Importance of 2D Representation, Positions, and Objects},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=0gOQeSHNX1},\nnote={under review}\n}" }, "abstract": { "value": "The Abstraction and Reasoning Corpus (ARC) is a popular benchmark focused on *visual reasoning* in the evaluation of Artificial Intelligence systems. In its original framing, an ARC task requires solving a program synthesis problem over small 2D images using a few input-output training pairs. In this work, we adopt the recently popular *data-driven* approach to the ARC and ask whether a Vision Transformer (ViT) can learn the implicit mapping, from input image to output image, that underlies the task. We show that a ViT—otherwise a state-of-the-art model for images—fails dramatically on most ARC tasks even when trained on one million examples per task. This points to an inherent representational deficiency of the ViT architecture that makes it incapable of uncovering the simple structured mappings underlying the ARC tasks. Building on these insights, we propose ViTARC, a ViT-style architecture that unlocks some of the visual reasoning capabilities required by the ARC. Specifically, we use a pixel-level input representation, design a spatially-aware tokenization scheme, and introduce a novel object-based positional encoding that leverages automatic segmentation, among other enhancements. Our task-specific ViTARC models achieve a test solve rate close to 100% on more than half of the 400 public ARC tasks strictly through supervised learning from input-output grids. This calls attention to the importance of imbuing the powerful (Vision) Transformer with the correct inductive biases for abstract visual reasoning that are critical even when the training data is plentiful and the mapping is noise-free. Hence, ViTARC provides a strong foundation for future research in visual reasoning using transformer-based architectures." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Abstraction and Reasoning Corpus", "Abstract Visual Reasoning", "Transformers", "Vision Transformers" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/01c0fe21bad2c80ddf91ed7cd127e3d5d8f0deb7.pdf" }, "presentation": null, "primary_area": { "value": "generative models" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/392b7455561e15c58c5053d1ab3362104851a621.zip" }, "title": { "value": "Tackling the Abstraction and Reasoning Corpus with Vision Transformers: the Importance of 2D Representation, Positions, and Objects" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "Inovation and contribution:\n1. The core of this paper combines the scores from the Global Router and the Local Router, then performs a top-k operation based on the combined scores. a. Line 85: The descriptions of the algorithm in the preceding and following contexts are clearly conflicting, making it contradictory to describe the algorithm effectively. b. Relying on GPT-4 turbo to gather information from the full text means the process heavily depends on GPT-4's capabilities. Furthermore, as seen in line 346 and the subsequent ablation experiments, the value of $\\alpha$ is significant, indicating a high weight assigned to the Global score.\n2. The paper extensively references the work of Muqeeth but fails to explain the rationale behind these references, raising concerns about the originality of the article.\n3. Relying on GPT-4-turbo for global semantic information raises some questions about the overall workload and originality of the research. From the formula in line 346 and the subsequent ablation experiments, it appears that the global score heavily influences the final affinity score, making the expert selection largely dependent on GPT-4-turbo. Additionally, the potential latency issues introduced by using GPT-4-turbo during inference are not addressed, which is a significant concern. Overall, placing so much emphasis on GPT in the paper could weaken its persuasive impact.\n\nFormat and Typo Issues:\n1. 60) Formatting issue: The left parenthesis is missing.\n2. 67-68) The sentences are semantically repetitive.\n3. 140) What does the arrow ($\\rightarrow$) signify here?\n4. 216) This figure is drawn in a hard model; it’s unclear what it is trying to convey.\n5. 279) The explanation of $\\Psi$ is missing.\n6. 289-290) The notation used here is confsued.\n7. 301) There is no explanation for the newly introduced $\\sigma$.\n8. 319) Suddenly introducing the variable 𝑡 t from line 289 is confusing.\n9. 323) Why isn’t the normalization process presented in a formula? Section 4.2 is poorly written, relying solely on descriptive language, lacking organization.\n10. 354) The T0 held-in dataset is omitted." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The insights of the paper are to be praised.\n2. Very interesting topic and focus on the router optimization. \n3. Use the big model and small model together to solve the problem \n4. Experimental results are good." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper focuses on addressing the trade-off between performance improvement and generalization ability in expert modules. It assumes that this issue arises from the lack of global semantic context in token-level routing strategies, it then seeks to resolve this problem by combining global semantics with token-level information through the use of both global and local expert routers during routing." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Writing and Presentation:The paper could benefit from some polishing. There are a number of typos and semantic issues, and the overall formatting could be improved for better readability. Additionally, some figures are a bit challenging to interpret. For instance, Figure 1 is only referenced in Appendix B but appears as the first figure in the Related Works section, which can disrupt the flow and clarity for the reader.\n\n2. Clarity of Background and Concepts: The background and explanation of key concepts in the paper could be clearer. While there are many references to ideas and works by Yadaav, the connections and explanations aren’t sufficiently detailed, which may leave readers a bit confused. In my initial reading, I found myself questioning whether the discussion pertained to a Mixture of Experts (MoE) scenario or Model Merging.\n\n3. Logical Flow and Mathematical Details: The paper seems to lack some logical coherence, especially regarding mathematical descriptions and derivations. There are no thorough mathematical proofs provided, and the modeling of the scenario feels a bit scattered across different sections. Some variable explanations are incomplete, which can be frustrating. Moreover, discussions around problems and solutions could be more precise. For instance, when mentioning that routing strategy issues stem from a lack of global semantics, it would be helpful to have more rigorous mathematical reasoning or experimental evidence to support this claim.\n\n4. Inconsistencies: There are some notable inconsistencies in the paper. For example, in line 85, it mentions that the Global Router selects the top-2 experts based on global semantics. However, the description of the Global Router algorithm starting from line 329 doesn’t reference any top-2 (or top-k) selection process. The top-k expert selection is only brought up later around line 347, based on the final score calculated from the weighted sum of global and local affinity scores. Clarifying these points would enhance the overall coherence of the paper.\n\n5. The experimental design could use some improvement. The main experiments lack detailed explanations, and Figure 3 is somewhat unclear. Many of the experimental configurations seem to mirror those from Muqeeth's work without providing enough context, which might raise questions about the originality of this study. Additionally, the ablation experiments focus on relatively trivial variables, while more significant factors—such as the differences between excluding and including the global semantics generated by GPT-4-turbo—are overlooked. Addressing these points could enhance the depth and rigor of the research." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "* Could you explain why T5 was chosen as the primary architecture for evaluation? Have you conducted any preliminary experiments with decoder-only models e.g. (LLama3-8B etc)?\n* Could you provide more details about the evaluation metrics used in Table 1? What exactly do these numbers represent?\n* Since this design integrates routing into the LoRA inference procedure, could you provide a detailed analysis of the additional computational overhead? How much will the inference latency be affected by such design?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "* The paper leverages LLMs to generate semantic task descriptions, providing global context for routing decisions which is a unique approach not explored in previous routing methods\n* The paper well address the limitations of current approaches (focusing on either held-in or held-out tasks) and provides a novel solution integrating both." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents GLIDER, a method that combines global semantic and local token-level routing for LLMs. The key innovation is using an LLM to generate task instructions that guide expert model selection globally, while using token-level routing locally. Tested on T5-based models with T0 and FLAN benchmarks, GLIDER improves performance on held-in tasks by 6.6% while maintaining strong generalization on held-out tasks. This shows that incorporating LLM-driven semantic understanding helps achieve better routing compared to pure token-level approaches." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* The experiments focus solely on T5, an older encoder-decoder architecture. The effectiveness of GLIDER on modern decoder-only models (like GPT family, LLaMA, etc.) remains unproven, which is crucial given these are now the mainstream architectures for LLMs.\n* Table 1 lacks clarity on evaluation metrics and methodological details. Without clear metric definitions and evaluation protocols, it's difficult to fully assess and compare the reported improvements.\n* The routing design will bring extra computational overhead, how will GLIDER's inference latency change compare with the normal LoRA decoding methods (vLLM's lora inference module)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "# incremental addition of new experts\n\nOne question that is not addressed is the incremental addition of new experts/datasets pair, and what are the consequences (on normalization, etc.). It seems it should be trivial to do this incrementally, but a discussion in the appendix (and an example use-case when, say a previously held-out task becomes held-in so that the router now uses global instead of local weights) would certainly improve the quality of the paper\n\n# limit of global router\n\nthe global router is constructed using a sample of 3 questions, which may be ok for very simple and low-diversity datasets, but not for more complex and diverse tasks,. a more in-depth study of global router across a wider range of datasets with quantitative assessment of global router policies seems necessary" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "this work introduces a routing mechanism to tradeoff between local and global experts, to increase performance on held in tasks, without compromising capability to handle held out tasks.\n\nthe goal is clear and the approach is simple (as it is heuristic in nature)." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "the paper studies ensembles of LLM models (model MoErging), proposing a technique for selecting experts to route tokens to at global (for selection of experts for in profile tasks) and local levels (to have more flexibility to handle out of distribution tasks)\n\nthe paper is incremental in nature (with small differences with respect to Phatgoose, from architectural design choice, to experimental settings and way too many details, to the point it feels it should be named Phatgoose++ instead of Glider)\n\nalthough appealing, the approach is heurisitic in nature: given this, one would have expected a significantly larger experimental part, including a wider range of tasks (and possible comparison points beyond those adopted in Phatgoose), and a statistical relevant comparison of improvements" }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "# evaluation\nas this work has no theoretical basis, one would have expected a significantly larger experimental part to convince the reader of the generality of the approach on \n- a significant wider range of tasks (and possible comparison points beyond those adopted in Phatgoose),\n- further exhibiting a statistical relevant comparison of improvements\n\nthis is not the case, so the paper execution is far from being convincing.\n\nAdditionally, while the main advantage of this work is to increase performance of held-in tasks, authors additionally point out advantages that are too thin to be worth noting; and they do so in a disturbingly biased manner. For instance, authors claim 0.9% over held out tasks over Phatgoose (in bold), but the 0.9% (actually 0.88%) is the maximum observed across 5 held out datasets in Tab 3 (for the other 4 is 0.39%, 0.16% -0.53% and 0.3%) \n\n# empirical evidence of global router\n\nthe work is motivated to find *semantical* resemblance beyond tasks. however, the approach on held out tasks seems to leverage *syntactical* resemblance . Fig 1 shows held out tasks to systematically select two experts (one of which seems to be further common to a couple of tasks). Yet appendix B just shows the tasks to be syntactically similar: i..e, the Q&A pair has a cloze format, which is rather typical of simple benchmarks. As such, I am little reassured that the performance generalization will be maintained on more complex tasks, and this work is far from fully elucidating the robustness of the proposed method.\n\nQuantitative assessment of global experts over a wider range of diverse tasks (say few tens of datasets per type of answer) would have allowed to get true insights about the nature of global experts (e.g., whether the identified expert triggers \"cloze\" type answers, irrespectively of the semantic of the question)" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "My first question relates to the core motivation of improving held-in performance and the necessity of doing so given that for any held-in task, we always have access to the expert specialized to that task. Could the authors explain scenarios in which using GLIDER is preferable to simply selecting the specialized expert for a known held-in task? \n\nMy second question is to what extent is GLIDER more of a novel component to local routing schemes that aims to encode global context, as opposed to an entirely new model. If GLIDER is more a novel component than an entire model, then I think the authors should include ablation studies on the local router choice, in particular using Arrow." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1) \nThe core idea -- that incorporating global information of the specialization of finetuned expert models into local routing schemes can improve expert aggregation algorithms -- is intuitive and persuasive.\n\n\n2) \nThe use of an LLM to encode global semantic information of the overall expert specialization is a creative method for effectively integrating the required global context" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors propose GLIDER, which is a token-level and module-level hierarchical routing strategy for combining pools of parameter-efficient finetuned expert models, incorporating both local and global information over the specialization of the expert models. The local component learns a gating between LoRA modules which selects the best layer module from the pool of experts for each token. The global component uses an LLM to encode the overall task expertise of each expert, which is then incorporated into the routing scheme to enhance the routing such that it is sensitive both to local module-wise expertise and overall global expertise of the aggregated models. This scheme hopes to maintain strong generalizability to unseen tasks without sacrificing performance on held-in tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1) \nMy first concern relates to the overall problem setting of the paper and its core motivation of improving performance on held-in tasks. The authors claim that existing MoErging methods often prioritize generalization to unseen tasks at the expense of performance on held-in tasks, and indeed in Table 1 the authors report as one of their main results that GLIDER significantly outperforms baselines on held-in tasks. However, performance on held-in tasks is deemed unimportant precisely because we already have access to the specialized expert trained on that exact held-in task, and so we can always retain performance on any given held-in task by simply using the expert specialized to that task. Indeed, this is the reason why Phatgoose does not report results for held-in tasks. For this reason, the 'Oracle Expert' recorded in Table 1 for held-in tasks is not an Oracle but an attainable result for which we always have access - it is just the expert specialized to that given task. \n\nSo in this sense, given that for held-in tasks GLIDER still underperforms the expert specialized to that given task, I'm not yet convinced of one the proposed main benefits of GLIDER, since for any given held-in task we could easily get better performance by just selecting the corresponding expert specialized to that task. \n\nFundamentally, I'm not yet persuaded that the performance gains on held-in tasks justify the claim that GLIDER is an overall superior model, since by the problem setting these are not tasks we need to optimize for. If the authors can provide justification for why performance on held-in tasks is indeed important and why selecting GLIDER over the corresponding specialized expert for a given held-in task would be preferable, then I would be happy to change my score, but as it stands I'm concerned that a large proportion of GLIDER's performance gains are on tasks that we need not optimize for.\n\n2) \nA second concern is that GLIDER's architecture, by the authors' own acknowledgement, is basically identical to Phatgoose for the local component of the hierarchical router. This being the case, the contribution of the paper is more so appending a global context component to Phatgoose, rather than an entirely new model. It would therefore be informative to consider alternative backbones for the local component of the router, for example Arrow. This could help isolate the contribution of the global routing component and help to demonstrate the robustness of potential improvements brought about by the proposed inclusion of global information.\n\n3) \nSome grammar / spelling related issues:\n\nLines 53-54: 'However, MoE methods that train experts from scratch while MoErging utilizes a decentralized...' -> delete 'that'\n\nLine 72: 'retrieve the correct expert for all token at every module' -> tokens should be plural\n\nLine 264: 'our goal is to build a MoErging method that dynamically routing queries' -> should be 'routes queries'\n\nLine 286: 'this work specifically focuses in LoRA' -> focuses 'on' LoRA\n\nLines 288-289 'Given the $t^{th}$ input token activation $u_i$ -> should be $u_t$ I'm assuming?\n\nLine 332: 'added before the model to independently process the fully query' -> process the 'full' query\n\nLine 348: 'the output of the module for token activation $u_i$ is computed as $Wu_i + \\sum_{k \\in \\xi_{top}} w_k * B_kA_ku_i$ -> It looks like you've forgotten to actually define $w_k$, I'm assuming it's the softmaxed affinity score, but you've left it undefined." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024glider,\ntitle={Glider: Global and Local Instruction-Driven Expert Router},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=0gVatTOgEv},\nnote={under review}\n}" }, "abstract": { "value": "The availability of performant pre-trained models has led to a proliferation of fine-tuned expert models that are specialized to a particular domain or task. This has enabled the creation of powerful and adaptive routing-based “Model MoErging\" methods with the goal of using expert modules to create an aggregate system with improved performance or generalization. However, existing MoErging methods often prioritize generalization to unseen tasks at the expense of performance on held-in tasks. This limitation adversely impacts practical applicability, as real-world deployments require robust performance across both known and novel tasks. We observe that current token-level routing mechanisms neglect the global semantic context of the input task. This token-wise independence hinders effective expert selection, particularly for held-in tasks, as routing decisions fail to incorporate the holistic semantic properties of the task. To address this, we propose a novel method, Global and Local Instruction Driven Expert Router (GLIDER) that integrates a multi-scale routing mechanism, encompassing a semantic global router and a learned local router. As recent LLMs demonstrate advanced reasoning capabilities for semantic-related contexts, the global router leverages this ability to enhance expert selection. By utilizing the input query and an LLM, the router generates semantic task instructions that guide the retrieval of the most relevant experts across all layers. This global guidance is complemented by a local router that facilitates token-level routing decisions within each module, enabling finer control and enhanced performance on unseen and challenging tasks. Our experiments using T5-based expert models for T0 and FLAN tasks demonstrate that GLIDER achieves substantially improved held-in performance while maintaining strong generalization on held-out tasks. Additionally, we perform ablations experiments to dive deeper into the components of GLIDER and plot routing distributions to show that GLIDER can effectively retrieve correct expert for held-in tasks while also demonstrating compositional capabilities for held-out tasks. Our experiments highlight the importance of our multi-scale routing that leverages LLM-driven semantic reasoning for MoErging methods. Our code is attached as supplementary material." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Parameter Efficient Fine-Tuning", "LoRA", "Cross-Task Generalization" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/f4f192ceabeb35c360c8f4ee9054561613663198.pdf" }, "presentation": null, "primary_area": { "value": "transfer learning, meta learning, and lifelong learning" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/cd603e8628610244550f0422dfc0bc414ac3019e.zip" }, "title": { "value": "Glider: Global and Local Instruction-Driven Expert Router" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "* is it really fair to say that your algorithm is more computationally tractable when it is based on MLE?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "* With the (unfortunate) exception of the introduction, I found the paper mostly well-written and clear.\n\n* The paper studies an interesting problem, proposes a natural solution, and proceeds to analyze said solution. While I am not familiar with the immediately preceding related work (in IV), this seems clean to me." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper studies a two stage procedure for regression in the scenario where the errors are not conditionally independent. They first learn a conditional density to make use of instrumental variables and consequently solve a square loss erm problem weighted by the learned conditional density. They show that this procedure attains mostly standard nonparametric rates." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* The organization of the paper is hard to follow and the introduction is way too terse. As someone well-versed in nonparametric statistics but not necessarily IV methods, I had to skip ahead to section 4 to really understand what was going on. Stating that you are trying to solve some fixed point equation in the introduction is not conducive to most people's understanding of the problem you are solving. \n\n* My overall feeling is that the result is somewhat incremental. To my understanding the main difficulty lies making standard guarantees for MLE in Hellinger^2 compatible with the square loss. I could not entirely follow why this is so challenging and would encourage the authors to further explain why this is the case (for instance, in the very last paragraph of section 1, you mention this difficulty but do not really expand on it, nor do you reference the lemmata which might be useful for understanding this difficulty)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "please see above" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "RDIV offers several key advantages over existing methods. It addresses three significant limitations of prior literature: it eliminates the need for unique IV regression identification, avoids reliance on the often unstable minimax computation oracle, and supports model selection procedures." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper addresses the problem of nonparametric instrumental variable (IV) regression, a framework with wide applications across fields such as causal inference, handling missing data, and reinforcement learning. The objective in IV regression is to solve the conditional moment equation, 𝐸 [ 𝑌 − ℎ ( 𝑋 ) ∣ 𝑍 ] = 0, where 𝑍 serves as the instrument. The authors introduce RDIV, a regularized variant of the DeepIV method, marking the first work to provide rigorous theoretical guarantees for DeepIV. RDIV enhances generalization by incorporating Tikhonov regularization. Methodologically, RDIV follows a two-stage approach. The first stage involves learning the conditional distribution of covariates, while the second stage refines the estimator by minimizing a Tikhonov-regularized loss function." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "It is unclear how the method compares for example to recently developed methods (see arxiv:2405.19463; to appear at NeurIPS 2024) that completely avoids minimax formulations, as well as avoiding the need for two-stage procedures." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "How is Tikhonov regularization related to a function space parametrized by the neural network? It seems not straightforward to relate it to weight decay.\n\nIs there a computational gain when minimax optimization is no longer needed?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper is well-written, and the results are motivated well. I didn't go through the proofs, but the explanations after each result are insightful, and ease the reading. The theoretical contribution is solid." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper studies the nonparametric instrumental variable regression with Tikhonov regularization (RDIV), and proves that RDIV allows model selection procedures and matches the SOTA convergence rate. I agree with the author's claim that this work is the first attempt to provide rigorous guarantees for DeepIV. With Tikhonov regularization, the model selection procedure achieves the oracle rate and iterative RDIV matches the SOTA rate." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The numerical experiments are only based on simulated data. It would be better to have some results from real data to demonstrate the strength of the proposal." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "* To convert the MLE guarantee into an $L_2$ guarantee, the authors assumed a minimum density on the conditional density. What’s the benefits/drawbacks compared with the conditional mean embedding based methods (although it also requires some assumptions like HS operators)." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "* The authors discussed lots of the aspects for non-parametric clearly." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This manuscript is basically a technical paper, discussed about two-stage non-parametric IV and model-selection in the second stage when equipped with an additional $L_2$ regularization." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* I feel the problem studied in this paper is with limited novelty. The transformation between one-stage and two-stage algorithms and analysis is in general only a technical problem and have been discussed in different places like DualIV, and the $L_2$ regularization itself makes the model-selection easier (with the strongly convexity)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024regularized,\ntitle={Regularized Deep{IV} with Model Selection},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=0gqCIaBRQ9},\nnote={under review}\n}" }, "abstract": { "value": "In this paper, we study nonparametric estimation of instrumental variable (IV) regressions. While recent advancements in machine learning have introduced flexible methods for IV estimation, they often encounter one or more of the following limitations: (1) restricting the IV regression to be uniquely identified; (2) requiring minimax computation oracle, which is highly unstable in practice; (3) absence of model selection procedure. In this paper, we analyze a Tikhonov-regularized variant of the seminal DeepIV method, called Regularized DeepIV (RDIV) regression, that can converge to the least-norm IV solution, and overcome all three limitations. RDIV consists of two stages: first, we learn the conditional distribution of covariates, and by utilizing the learned distribution, we learn the estimator by minimizing a Tikhonov-regularized loss function. We further show that RDIV allows model selection procedures that can achieve the oracle rates in the misspecified regime. When extended to an iterative estimator, we prove that RDIV matches the current state-of-the-art convergence rate. Furthermore, we conducted numerical experiments to justify the efficiency of RDIV empirically. Our results provide the first rigorous guarantees for the empirically well-established DeepIV method, showcasing the importance of regularization which was absent from the original work." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Nonparametric estimator", "instrumental variables", "model selection", "causal inference." ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/36421fa6453920abc8f29f0ef6d69cd1560c852a.pdf" }, "presentation": null, "primary_area": { "value": "learning theory" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Regularized DeepIV with Model Selection" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "-" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "I am happy to increase my score and support this paper with a high confidence if the authors can provide an extensive discussion during the rebuttal on the assumptions in Azizian et al. (2023a) . In particular, my two major questions are: can the authors be more precise in which cases their assumptions are weaker than the ones in Azizian et al. (2023a). In particular, can you give an example for a class of distributions that are covered by this paper but not by Azizian et al. (2023a)? Moreover, can the authors explain why the proof in Azizian et al. (2023a) breaks for your assumptions and why it is not trivial to extend the proof?\n\n\n\nSmaller question:\nIsn't assumption 3.1 (1) always true satisfied by w<=1. Is it possible that this is a typo?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper addresses an important problem is generalization bounds/theoretic ML. In particular:\n\n- The paper is well written and the results are nicely presented. \n- The proof sketch in Section 4 is excellent. It is very easy to follow and often neglected in these types of papers\n- The proof idea is smart, non-trivial and interesting." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents novel bounds on for the DRO loss using the Wasserstein distance. In particular, they address the question of finding the minimal $\\rho$ used by the empirical robust loss such that the loss is an upper bound for the actual population loss. The main challenge is to overcome the dependency on the distance between W(P_n, P) \\sim n^{1/d}. While this problem has been studied in the literature, and dimension free bounds exist, this paper presents a proof requiring weaker assumptions." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Given that this is a more traditional field, I would expect a clearer comparison with the existing works. While the authors do a very good job in presenting the proof idea, it is not so clear how the proof fundamentally differs from existing works." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- What are the practical implications of the generalization guarantees compared to Azizian et al. (2023a)? Can you provide some numerical results analogous to Appendix H of Azizian et al. (2023a)?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The generalization guarantees of this work do not rely on restrictive assumptions like smoothness compared to the previous work (Azizian et al. 2023a). \n- This paper is well-structured, and the theoretical results and proof sketches are clearly presented." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper provides exact generalization guarantees for Wasserstein Distributionally Robust Optimization (WDRO) for a wide variety of models with compactness and finite Dudley's entropy assumptions. The results apply to radius $\\rho$ scaling as $O(1/\\sqrt{n})$, which does not suffer from the curse of dimensionality. The results also cover the double regularization case." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- In Section 3.2, the authors discussed how their results on generalization guarantees apply to linear regression and logistic regression. However, more complicated models such as neural networks with ReLU or other smooth activation functions (e.g. GELU) are not discussed. \n- The theoretical results require a lower bound on $n$, while Theorem 3.4 of Azizian et al. (2023a) applies to all $n \\ge 1$. The implications of this requirement should be discussed." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "- Why not submit to JMLR? The paper is very rigorous and rather long and technical for an ML conference? You also examine the problem in good detail.\n\n- Could you provide a simple example in Theorem 3.2, where the optimal coupling is known under (say) Gaussianity assumptions?\n\n- I'm a bit confused. What does $\\operatorname{argmax}_{\\Xi}\\,f$ mean in (5) a sup norm or something? \n\n- Why is $\\min\\{ c(\\xi,\\zeta): ... \\}$ measurable? In particular, (independent of the meaning of the argmax, above question), why is there a measurable selection $\\xi\\mapsto \\zeta$? Without this, its not clear that $\\rho_{\\operatorname{crit}}$ is well defined. I'm guessing this is Berge's theorem (which is in Aliprantis & Border) somehow, but please spell it out for us :)\n\n- Each result assumes that the (difficult to interpret) $\\rho_{\\operatorname{crit}}$ is \"large enough\". Can you please provide a general set of conditions ensuring that $\\rho_{\\operatorname{crit}}$ can be bounded away from $0$. \n\n- Is it fair to compare, verbally, our results to those of Fournier et al. (and similar bounds, say, found in [1])? Since you are considering a small ball around the empirical measure while their results guarantee a minimal radius such that the empirical measure contains the true measure whp. Furthermore, those rates are only tight (afaik) when the measure is very spread out; more precisely, it is Alhors $d$-regular, see e.g. [3] for a nice clean proof. \n\n- In theorem 3.2, why is $\\pi^{P,Q}\\ll \\pi_0$? To be this isn't directly evident... I.e.\\ why is the RHS not trivially $-\\infty$ in general?\n\n\n\n[1] Graf, Siegfried, and Harald Luschgy. Foundations of quantization for probability distributions. Springer Science & Business Media, 2000.\n[2] Otto, Felix, and Cédric Villani. \"Generalization of an inequality by Talagrand and links with the logarithmic Sobolev inequality.\" Journal of Functional Analysis 173.2 (2000): 361-400.\n[3] Kloeckner, Benoit. \"Approximation by finitely supported measures.\" ESAIM: Control, Optimisation and Calculus of Variations 18.2 (2012): 343-359." }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper is well written, interesting, and theoretical and provides very nice lower bounds on the robust empirical risk. The results are nice, and so is the use of set-valued analysis to derive them. Several relevant examples are considered, making a large portion of how these results can be used nearly transparent." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The proposed paper provides lower bounds on the robust empirical risk under unorthodox but interesting scaling limits on the radius of the Wasserstein ball around the empirical risk. The paper uses some cool techniques which are not often seen in machine learning. \n\nThe paper is relatively clearly written. However, I think there are a few little things here and there which are either difficult to justify (in the current form) or perhaps not well-defined (see below). Also, the introduction is excessively general while the setting rapidly collapses to a much more specific setting shortly after." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Nevertheless, I think some of the assumptions are a bit opaque (see below), and I'm not certain some quantities are well-defined.\n\n*Minor*\n- Citing Cuturi and Perée's book is odd when mentioning the Wasserstein distance. Perhaps the original source or a book on optimal transport such as Villani's book, would be more natural, IMO.\n\n- The definition of Wasserstein distance circa (1) is incorrect, $\\Xi$ must be Polish, and c *must* be a power $p\\in [1,\\infty)$ of a *metric* topologizing $\\Xi$; what you write is just some transport problem. Eg if c is not symmetric, then $W_c$ is not a metric in general, or if $c(x,y)=0$ for all $x,y$ then $W_c$ cannot separate points.\n\n- Perhaps \"suitable\" distributions is more appropriate before (1), since the distance explodes if these have no finite moment. \n\n- Line 65: bad grammar: \"it does not introduce approximate term\" also imprecise.\n\n- Line 66: Is Wainright's book and Boucheron the best reference? Perhaps older papers, e.g. on VC dimension, Bartlett's old papers on Rademacher complexity, or old papers on chaining are more natural references?\n\n- Line 69: \"This theoretical feature is specific to WDRO and highlights its potential to give more resilient models.\" This can be **much** less hand-wavy. Please explain more precisely/mathematically.\n\n- Line 149: In a metric space $(x,..)$ not \"In (X,..) a metric space\".\n\n- Assumption 1 vs. Line 145: You say that $\\Xi$ is just a measurable space, then later you say its a compact metric space. Why no be forthright and say its a metric space on line 145. Similarly, why is $\\mathcal{F}$ an arbitrary family of functions, then straightaway after is actually a compact set of continuous functions.\n\n- Line 176: Why jointly Lipschitz? If $\\Theta$ is compact, then since you already assumed $\\Xi$ is compact, then it is enough for $\\Theta\\times\\Xi\\ni (\\theta,\\xi)\\mapsto f(\\theta,\\xi)\\in \\mathbb{R}$ to be continuous; to deduce the compactness of $\\{f(\\theta,\\cdot):\\,\\theta \\in \\Theta\\}$ by the currying Lemma. \n\n- Line 176: Not sure why you say \"if $\\Xi$ is compact, since this was assumed a few lines earlier on the same page.\n\n- Maybe more natural examples come from Arzela-Ascoli...\n\n- Should the definition of the Dudley entropy integral really be in a footnote, while more basic ideas are in the main text.\n\n- Line 221: The words \"the metric\" are missing.\n\n- Line 223: There are many more references of the use of this type of metric, especially in exponential convergence rate results for Markov chains (wrt $W_1$ over countable metric spaces with this distance).\n\n- Line 245: \"sample randomness\" (I know what you mean...but the word independent is misleading as this \n\n- Assumption 1: Why call (2) jointly continuous, it is just standard continuity (actually inform continuity by compactness)." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "Exact generalization guarantees for Wasserstein distributionally robust models with dimension-free sample rates." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024universal,\ntitle={Universal generalization guarantees for Wasserstein distributionally robust models},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=0h6v4SpLCY},\nnote={under review}\n}" }, "abstract": { "value": "Distributionally robust optimization has emerged as an attractive way to train robust machine learning models, capturing data uncertainty and distribution shifts. Recent statistical analyses have proved that generalization guarantees of robust models based on the Wasserstein distance have generalization guarantees that do not suffer from the curse of dimensionality. However, these results are either approximate, obtained in specific cases, or based on assumptions difficult to verify in practice. In contrast, we establish exact generalization guarantees that cover a wide range of cases, with arbitrary transport costs and parametric loss functions, including deep learning objectives with nonsmooth activations. We complete our analysis with an excess bound on the robust objective and an extension to Wasserstein robust models with entropic regularizations." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "generalization guarantees", "optimal transport", "distributionally robust optimization", "nonsmooth analysis" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/54769a76e1c4dcf908ecc8ba3447d3eadf3d62fc.pdf" }, "presentation": null, "primary_area": { "value": "optimization" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Universal generalization guarantees for Wasserstein distributionally robust models" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. Could you provide justifications for the difference between your optimizer prompt and OPRO’s meta-prompt [1] for prompt optimization? \n\n[1] Yang, C., Wang, X., Lu, Y., Liu, H., Le, Q. V., Zhou, D., & Chen, X. (2023). Large Language Models as Optimizers (No. arXiv:2309.03409). arXiv. http://arxiv.org/abs/2309.03409\n\n[2] Pryzant, R., Iter, D., Li, J., Lee, Y. T., Zhu, C., & Zeng, M. (2023). Automatic Prompt Optimization with “Gradient Descent” and Beam Search (No. arXiv:2305.03495). arXiv. http://arxiv.org/abs/2305.03495" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The core idea addresses a critical limitation of iterative/reflective methods that only focus on immediate feedback. \n\n2. This work covers a wide array of recent literature, and the presentation is clear." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Existing automatic optimization methods only focus on immediate feedback, which can be easily trapped by the local optima. HessianGrad is analogous to second-order derivative methods by taking into account the feedback over multiple iterations. Experimental results show consistent gains of HessianGrad in prompt optimization, solution refinement, and code optimization." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Besides the analogy provided in Eq. 2 and Eq. 3, the core implementation of HessianGrad is to include a meta-prompt that encourages LLM to reflect over multiple turns as shown in Pg. 12. However, this is arguably analogous to the real Hessian matrix that the work is trying to deliver. Moreover, whether LLM is able to capture the second-order phenomenon is also questionable and lacks justification in this work. \n\n2. The actual technical contribution of this work is to provide a more refined meta-prompt to the original TextGrad’s meta-prompt. First, the contribution of the refined meta-prompt appears to be limited. Second, the OPRO work [1] has also included similar meta-prompts to reflect over multiple turns by feeding the iterative optimization trajectory into the context of LLM. Therefore, the novelty of this work is limited and appears more as an incremental improvement on TextGrad.\n\n3. The selected baselines in main experiments are also questionable. Most baselines are TextGrad and M-TextGrad. However, for each task, competitive baselines, e.g. ProTeGi [2] in prompt optimization, are not compared." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. In general computational frameworks, HessianGrad is computed by adding a small quantity in the iteration direction and recalculating the gradient once, followed by taking the finite difference of the gradients. From a practical standpoint, can the direct finite difference version of similarity $\\mathcal{S}(r(p_t), r(p_{t-1})) = \\frac{|| \\mathcal{L}(r(p_t)) - \\mathcal{L}(r(p_{t-1}))||}{||p_t - p_{t-1}||}$ save computational costs (since $\\mathcal{L}(r(p_{t-1}))$ and $p_{t-1}$ are both values computed from the previous iteration) and achieve similar effects?\n2. There is a typo in Equation (3) in Section 3. The second-order partial derivative should be denoted as $\\frac{\\overset{\\sim}{\\partial}^2\\mathcal{L}}{\\overset{\\sim}{\\partial}p_t^2}$ rather than $\\frac{\\overset{\\sim}{\\partial}^2\\mathcal{L}}{\\overset{\\sim}{\\partial}^2p_t}$." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "This paper is well written with clear presentation of the new algorithm, and the experiments are rigorous with repeatability." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a new method for optimizing system prompts through gradient descent. The authors outline the limitations of the classical method \"TextGrad\", and present solutions to mitigate these issues. The second-order derivative (i.e., HessianGrad) is introduced into \"TextGrad\", thereby reducing the likelihood of the system prompt getting trapped in local minima. The authors conduct empirical experiments on three tasks: prompt optimization, solution optimization, and code optimization. The conclusion is that the proposed new method outperforms the naive TextGrad method and the Momentum-Enhanced TextGrad method across four mainstream models." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Overall, this paper may have an algorithmic contribution, but supplemental details are required on both theoretical and experimental aspects.\n\n1. Theoretical aspect:\n- When defining the similarity $\\mathcal{S}$ of the prompt before and after each iteration, you do not multiply it by any hyperparameters and directly add it to the original TextGrad to obtain the HessianGrad. Consider adding two hyperparameters, $\\beta_1$ and $\\beta_2$, to the two terms of HessianGrad, similar to Adamw. Could you conduct an ablation study on the effect of adding these hyperparameters, or provide justification for why they were not included in the current formulation?\n- You use the $L_1$ norm to define $\\mathcal{S}$, however, when measuring semantic similarity, cosine similarity is more commonly used as it signifies that the loss $\\mathcal{L}$ before and after iteration is closer in direction, while $L_1$ primarily signifies that $\\mathcal{L}$ is closer in numerical value. It is recommended to provide a rationale for choosing the $L_1$ norm as a similarity metric. (Simple reasons are acceptable, such as \"$L_1$ norm is easier to compute\"). Could you compare the performance of your method using L1 norm versus cosine similarity, or provide empirical justification for why L1 norm was chosen over other common similarity metrics?\n\n2. Experimental aspect:\n- Please provide the loss curves for HessianGrad, M-TextGrad, TextGrad, and CoT of their iterative processes for a representative example from each of the three tasks (prompt optimization, solution optimization, and code optimization) to demonstrate the effect of \"HessianGrad Escaping Local Optima\" as shown in Figure 1.\n- Calculating HessianGrad typically requires more computational resources. Could you provide a detailed comparison of computational resources (GPU memory, runtime) for HessianGrad versus the baseline methods across all three tasks." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please respond to the concerns in the \"Weaknesses\" part.\n\nQ1. The motivation (line 223~226) of introducing the similarity function does not match well with the second-order optimization theory. How can the similarity of the responses provide second-order information?\n\nQ2. The concrete definition of the similarity function on line 244 is meant to connect with the formulation of second-order optimization. However, this definition is contrary to the motivation in line 226 (\"more gradual and thoughtful evolution of the response over multiple iterations\"), since larger similarity means larger fluctuation between successive steps, according to this definition. Moreover, this definition is actually focusing on changes in feedback ($L(r(p_t))$) instead of response ($r(p_t))$), and this is a point that contradicts the motivation of this paper. Please clarify how the definition aligns with the stated motivation." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The paper writing is clear and easy to understand.\n2. The proposed optimization method achieves considerable improvements on a variety of tasks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a new LLM-based textual optimization method that takes the evolution of LLM systems responses across iterations into account. Improvements are achieved on multiple textual optimization tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Although the authors have written about the difference between momentum-based methods and HessianGrad, the novelty of transferring the focus from feedback similarity to response similarity is somewhat weak. The authors should include more convincing ablation experiments to verify that tracking the dynamics of responses is more effective.\n\n2. The second-order Hessian formulation can not provide sufficient theoretical support for the optimization framework. The relationship between tracking feedback and tracking responses is not comparable to first-order and second-order optimization." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024hessiangrad,\ntitle={HessianGrad: Optimizing {AI} Systems with Hessian-Aware Textual Gradients},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=0hc7iQLhCt},\nnote={under review}\n}" }, "abstract": { "value": "Recent advancements in large language models (LLMs) have significantly enhanced the ability of LLM-based systems to perform complex tasks through natural language processing and tool interaction. However, optimizing these LLM-based systems for specific tasks remains challenging, often requiring manual interventions like prompt engineering and hyperparameter tuning. Existing automatic optimization methods, such as textual feedback-based techniques (e.g., TextGrad), tend to focus on immediate feedback, analogous to using first-order derivatives in traditional numerical gradient descent. However, relying solely on first-order derivatives can be limited when the gradient is either very small or fluctuates irregularly, which may slow down or stall optimization. To address these limitations, better adaptation in regions with small or fluctuating gradients is necessary. Second-order gradient methods, which incorporate the Hessian matrix, offer a promising solution by enabling more precise adjustments. Inspired by this, in this paper, we introduce HessianGrad, a novel optimization method that leverages textual feedback and tracks the iterative evolution of LLM systems responses across iterations, leading to more dynamic and adaptive optimization. We evaluate the effectiveness of HessianGrad on three tasks: prompt optimization, solution optimization, and code optimization. Experimental results demonstrate that HessianGrad consistently improves performance across all three tasks, achieving a **7.8%** improvement in prompt optimization, a **20.72%** gain in solution refinement, and a **29.17%** increase in code optimization compared to baselines, highlighting its adaptability and effectiveness in optimizing LLM-based systems." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "LLM", "Prompt Optimization", "Gradient Descent" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/f05fc8da8a3b631d67a8b72c55683865f055e7e6.pdf" }, "presentation": null, "primary_area": { "value": "foundation or frontier models, including LLMs" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "HessianGrad: Optimizing AI Systems with Hessian-Aware Textual Gradients" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "- Does the test time objective result in networks that are idempotent on OOD samples? Presumably once the function drifts from the fixed anchor function that test time loss no longer reflects idempotence? It would be good to see measures of idempotence on training and test samples.\n\n- How does the intuition about idempotence being a generalization of orthogonal projection hold up? The proposed method considers idempotence only in the y-variable, not the entire function." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- Idempotent training is an interesting and novel approach for addressing the out-of-distribution generalization problem.\n\n- The paper shows successful application of the considered approach on a large number of experimental settings, including in online-learning settings." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "- This paper proposes a test-time-training based approach to address the distribution shift or OOD generalization problem by learning models that are idempotent.\n\n- In particular, this paper proposes a method where models $f_{\\theta}: \\mathcal{X} \\times \\mathcal{Y} \\to \\mathcal{Y}$ are trained by minimizing both the difference between $f_{\\theta}(x, y)$ and $y$ as well as the difference between $f_{\\theta}(x, 0)$ and $y$, resulting in a model that is idempotent on the training set. Then at test time, models are optimized on the test data before running inference to make them idempotent on test inputs, by minimizing the difference between $f_{\\theta}(x, 0)$ and $f_{\\theta}(x, f_{\\theta}(x, 0))$ for an OOD input $x$.\n\n- The paper shows empirically for several different settings that idempotent test-time-training improves classification accuracy on out-of-distribution samples." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Test-time training requires undesirable extra compute for test-time optimization of the whole model. How much additional compute is needed compared to running inference on the base model? How does this method scale with model size?\n\n- The paper lacks analysis that explains how or why idempotent training is expected to improve out-of-distribution analysis. Further investigation and ablations that provide intuition for how this method works would be valuable.\n\n- The proposed method has not been evaluated on standard real-world out-of-distribution generalization benchmarks such as DomainBed [Gulrajani and Lopez-Paz 2020] and WILDS [Koh et al 2020]. The presented experiments are on smaller models/datasets.\n\nReferences:\n\nGulrajani, Ishaan, and David Lopez-Paz. \"In search of lost domain generalization.\" arXiv preprint arXiv:2007.01434 (2020).\n\nKoh, Pang Wei, et al. \"Wilds: A benchmark of in-the-wild distribution shifts.\" International conference on machine learning. PMLR, 2021." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. Do you use a special encoding or placeholder value instead of 0 for regression tasks?\n2. If the purpose of the neural \"don't know\" zero is purely for contrast and its specific value is not important, will it be more computationally efficient to use a representative value such as the median of $y_i$ where $i \\in \\text{training set}$?\n3. In EMA (Morales-Brotons et al., 2024), when $\\alpha = 1$, the online $IT^3$ aligns with the offline version, and when $\\alpha = 0$, it encounters the collapse issue described in Section 3.2. Could you provide guidance on selecting the value of $\\alpha$ or share experimental results demonstrating performance across different $\\alpha$ values?\n4. The right panel of Figure 2 and Figure 16 both appear to represent the car data. Could there be a potential mistake or duplication here?\n\nMorales-Brotons, D., Vogels, T., & Hendrikx, H. (2024). Exponential moving average of weights in deep learning: Dynamics and benefits. Transactions on Machine Learning Research." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The authors extend the concept of TTT (Sun et al., 2020) by incorporating idempotence, offering a simple yet elegant solution. The approach is intuitive and Figure 1 is particularly helpful for quickly grasping the core idea, even for those who are not experts in distribution shift." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces $IT^3$, a generic method for addressing distribution shift challenges. By enforcing idempotence, $IT^3$ sequentially adapts to potential distribution shifts during inference. The authors show the performance of $IT^3$ across diverse scenarios." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. (cf Figure 1) During training, in addition to feeding the standard $(x, y)$ pair to train the model, the authors also input the $(x, 0)$ pair to ensure the model satisfies the property of idempotence, referring to the zero as a neutral “don't know” signal. While this approach may work in classification tasks where zero could represent a new “don’t know” class, in regression tasks, it is unclear how the model differentiates this zero from an actual label of 0.\n2. (Line 211) Online $IT^3$ appears to rely significantly on the use of Exponential Moving Average (EMA). However, the authors did not provide a citation for this technique. \n3. In the first experiment (Section 4.1 on tabular data), the method of randomly substituting features with zeros in the test set may resemble a missing data scenario rather than a true distribution shift. In other experiments, the authors simulate distribution shift by splitting the training and testing data based on a specific covariate to create distinct distributions. It is unclear why the same method was not applied for the first experiment. If the authors prefer using the zeroing method, they should include figures or statistical tests to substantiate that the training and testing data are distributionally different, rather than relying solely on intuition.\n4. There are unnecessary abbreviations throughout the manuscript. For instance, “such that” is shortened to “s.t.” in line 490. The proposed method, $IT^3$, is inconsistently referred to as ITTT in parts of the manuscript, such as in Figure 13.\n5. Figures 14 to 16 are not mentioned or referred to in the main text. This omission is unusual and may confuse readers as to the purpose or relevance of these figures.\n6. Figures 12 and 15 have the exact same title.\n\nSome minor improvements and spelling corrections for clarity:\n1. (Line 22) $x$ (not $\\mathbf{x}$) is not mentioned before.\n2. (Line 77) $y_2$ is not mentioned before.\n3. (Line 85) \"th input\" should be corrected to \"the input\".\n4. (Line 131) Missing right parenthesis: $f(f(z))$.\n5. (Line 490) “s.t. that” should be replaced with “such that”.\n6. (Line 495) Remove the extra colon (\":\").\n7. In Table 1, the title references “qualitative” results, but the data presented are numerical and should be described as “quantitative” results." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "As mentioned before, I find this work quite interesting but the lack of proper baselines push me towards a weak reject opinion. As for specific questions and ways that the authors can improve their work:\n* More baselines are needed on all tasks; even if the method does not translate exactly to the setup considered, the authors could perform minor adaptations so that it does. For example, why not consider additional self-supervised tasks as discussed at Sun et al. (2020)? In the case where simple rotation prediction might not apply, something simple like denoising an artificially corrupted image could still work as a self-supervised task. Another example would be on the online setup; there, methods from the TTA literature could be applied, such as the work of [1] which works even on the instance level and without batch normalisation.\n* Apart from the CIFAR-C case, most other distribution shifts are generated by just partitioning the datasets according to some rule and then training on a subset while considering the other as OOD. This is a bit of a constrained setting and I would encourage the authors to consider more diverse shifts, as that would better highlight the usefulness of IT$^3$. For example, why not add noise to the road segmentation task, in the same way that it was done for CIFAR 10? This could be plausible real world setting where there is a fault in the sensor, thus the images come out corrupted. \n* How is the label encoded in the input in the various settings considered? This is an important information for reproducibility. Furthermore, is the loss at Eq. 2 used for all settings (even when it might not make much sense, such as classification)? \n* In the online setting, the authors consider a smooth transition between the distribution shifts, which might not be practically realistic. How does the method behave when the transition between distribution shifts is not smooth? \n* How many update steps on each datapoint do the authors do? Does the test time optimization reach a fully idempotent model and does “more idempotence” provide better performance? \n\n\n[1] Towards Stable Test-Time Adaptation in Dynamic Wild World, Niu et al., ICLR 2023" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "* The paper is mostly well written; the ideas are explained clearly and reasonable intuitions are being given. Same goes for the experimental evaluation and discussion of the settings. \n* The idea is simple and I find it quite interesting. While the main architecture of the method relies heavily on the prior work Durasov et al., the application on the TTT setting is novel, since, as far as I know, idempotence hasn’t been explored in such a scenario. \n* The tasks considered in the experiments are quite diverse, which is nice to see. They range from simple tabular data, to standard image classification, to regression and dense prediction with various architectures. \n* The authors demonstrate improvements with IT$^3$ on all settings considered upon (admittedly very simple) baselines." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work presents IT$^3$ a method for test time training that relies on the concept of idempotence. During training, the authors train the neural network to be able to predict the ground truth label by either conditioning on it (by concatenating it to the input) or by assuming a placeholder value for it (by concatenating a $\\mathbf{0}$ value to the input). At test time, the authors then fine-tune the model on unlabelled data by matching the predictions of the model when using $\\mathbf{0}$ as the label with that of the model when conditioning on its own output at the $\\mathbf{0}$ label. This essentially leads to a soft idempotence constraint which, according to the authors, allows the model to move out-of-distribution data closer to in-distribution ones and thus improve performance when distribution shifts happen at test time. The authors then evaluate IT$^3$ on a variety of tasks and architectures." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* The main weakness I see in this work is the (almost complete) lack of proper baselines. For example, on only the CIFAR task there is another TTT baseline but on all of the others the authors just compare against not doing anything. This make it hard to gauge significance of the method against other prior art.\n* The work makes claims about how idempotence can be seen as a generalisation of a projection and that it allows to map the OOD data to the in-distribution. Thus, while the authors do spend some time to explain why would intuitively their method works, they do not have any ablation studies to verify that these exact intuitions hold in practice. \n* IT$^3$ as a method requires two sequential forward passes through the model, so in practice it can be slow and the authors do not discuss the efficiency of IT$^3$ relative to other works." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please refer to the weaknesses section (especially point 2). I am happy to raise my score if my concerns are resolved." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The main strengths of this paper:\n\n1- Simplicity of the approach: the method is easy to understand and to implement. \n\n2- The paper is generally well-written and easy to follow.\n\n3- The breadth of the experiments: The authors are commended for the variety of different tasks that they tested their proposed methods on." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a new method for test-time training: at training time the model is trained to be Idempotent, and at test-time two copies of the models are leveraged (one frozen and one adapted) to encourage the model to be idempotent under distribution shifts. Experiments are carried out on different tasks to demonstrate how versatile the proposed method is." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The main weaknesses of this paper are:\n\n1- The method section, while simple, is a bit confusing. Why does the paper present two variants of IT$^3$? The experiments in section 4.6 and Table 1 clearly favors the online version of the method over the one with frozen predictor. This questions why not leveraging the online version in all the experiments? Is there an experimental setup where the offline one is much better than the online one? Why doesn't the paper present one method and treat the other as a special case/variant that is less powerful?\n\n2- The main weakness of this paper the lack of baselines in the experimental section. In all of the presented results, a very suboptimal version of TTT is compared against in **one experiment** only. This really questions how strong is IT$^3$/IT$^3$-online is when compared against strong baselines. Here are some suggestions for necessary experiments:\n\n2a. Since TTT/TTT++[A] are directly comparable with IT$^3$, then I suggest to *at least* have them in all of the experiments with one-to-one comparison in terms of batch size and model architecture. Also, consider adding the performance of IT$^3$ under batch size=1 in Figure 3 results.\n\n2b. Another strong baseline that is suitable for both classification and regression tasks is ActMad [B]. A direct comparison against this baseline is also necessary is all the presented experiments.\n\n2c. Another line of work that is directly comparable is Test-Time Adaptation (TTA). TTA works under a more conservative setup where no control over the training process is assumed. It is also important to compare against the current SOTA TTA method EATA [C] or more closely the dataset distillation method from [D] to further demonstrate the superiority if the proposed method.\n\n2d. Since this is a 'test-time' method, a discussion on its computational requirements is necessary. How would the performance be when evaluated under computational constraints [E]?\n\n2e. Benchmarks used in this work are somewhat small scale. Experiments on larger benchmarks such as ImageNet-C [F] and ImageNet-3DCC [G] in the classification setting are necessary. Similar arguments follow for regression tasks where for example one can follow the object detection experiments from ActMAD. \n\n3- In section 4.1, I am not sure about the Distribution Shift introduced in this experiment. For instance, the performance of non-optimized does not constantly degrade. Why zeroing out features is a good way of modeling distribution shift instead of adding random noise to the features? Can you please comment on this and provide justification for their choice of distribution shift.\n\n4- Missing references: [B, C, D, E, F, G].\n\n[A] TTT++: When Does Self-Supervised Test-Time Training Fail or Thrive?, NeurIPS 2021\n\n[B] ActMAD: Activation Matching to Align Distributions for Test-Time-Training, CVPR 2023\n\n[C] Efficient Test-Time Model Adaptation without Forgetting, ICML 2022\n\n[D] Leveraging Proxy of Training Data for Test-Time Adaptation, ICML 2023\n\n[E] Evaluation of Test-Time Adaptation Under Computational Time Constraints, ICML 2024\n\n[F] Benchmarking Neural Network Robustness to Common Corruptions and Perturbations, ICLR 2019\n\n[G] 3D Common Corruptions and Data Augmentation, CVPR 2022" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "Adapting to distribution shifts at test time by training the model to be idempotent." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024it,\ntitle={{IT}\\${\\textasciicircum}3\\$: Idempotent Test-Time Training},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=0hyShAPeBj},\nnote={under review}\n}" }, "abstract": { "value": "This paper introduces Idempotent Test-Time Training (IT$^3$),\na novel approach to addressing the challenge of distribution shift.\nWhile supervised-learning methods assume matching train and test distributions, this is rarely the case for machine learning systems deployed in the real world.\nTest-Time Training (TTT) approaches address this by adapting models during inference, but they are limited by a domain specific auxiliary task. IT$^3$ is based on the universal property of idempotence. An idempotent operator is one that can be applied sequentially without changing the result beyond the initial application, namely $f(f(x))=f(x)$. \nAn idempotent operator is one that can be applied sequentially without changing the result beyond the initial application, that is $f(f(X)=f(X)$.\nAt training, the model receives an input $X$ along with another signal that can either be the ground truth label $y$ or a neutral \"don't know\" signal $\\mathbf{0}$. At test time, the additional signal can only be $\\mathbf{0}$. When sequentially applying the model, first predicting $y_0 = f(X, \\mathbf{0})$ and then $y_1 = f(X, y_0)$, the distance between $y_1$ and $y_2$ measures certainty and indicates out-of-distribution input $x$ if high.\n We use this distance, that can be expressed as $||f(X, f(X, \\mathbf{0})) - f(x, \\mathbf{0})||$ as our TTT loss during inference. By carefully optimizing this objective, we effectively train $f(X,\\cdot)$ to be idempotent, projecting the internal representation of the input onto the training distribution.\nWe demonstrate the versatility of our approach across various tasks,\nincluding corrupted image classification, aerodynamic predictions,\ntabular data with missing information, and large-scale aerial photo segmentation. Moreover, these tasks span different architectures such as MLPs, CNNs, and GNNs." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "idempotence;generalization" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/1ffbba0b731bde24db6dad11af404803012f7da4.pdf" }, "presentation": null, "primary_area": { "value": "unsupervised, self-supervised, semi-supervised, and supervised representation learning" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "IT$^3$: Idempotent Test-Time Training" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Can the authors clarify the questions above? Specifically, the extent to which DINO already offers certain degree of disentanglement and how the factors of variation of interest could be identified directly from these representations." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "* The paper evaluates the recent ideas of disentangled representation learning using weak supervision in a more realistic application.\n* The paper also presents an alternative to learning the disentangled representation from RGB images based on models pretrained at large scale.\n* The paper proposes a new sprites dataset to facilitate the interpretation of microscopy images." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents a study of disentangled representation learning on three microscopy image datasets. The representation learning strategy starts by training an Ada-GVAE model using a Textures-dSprites dataset introduced in this work. The dataset is supposed to reflect simple textures that could help interpret information in microscopy images. After training this model in a weakly supervised way, it is used to encode images of another domain, with optional unsupervised finetuning using a beta-VAE. The resulting features are low dimensional, and interpretable, and are used to train classifiers.\n\nThe ideas and the study are generally interesting, but the paper lacks technical novelty, is limited to a small-scale empirical evaluation only, and the experiments are incomplete to fully understand the value of the proposed strategy." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* The technical contribution is limited. Beyond the sprites dataset and the use of pretrained features, many of the ideas have been presented in previous works.\n* The experimental evaluation is limited to quantifying the impact of classifier types (GBT vs MLP) and input type (RGB vs DINO features). Many questions remain open regarding how much classification accuracy could be obtained without the proposed disentanglement procedure. Can the authors compare results of training a classifier directly with RGB images and another classifier with DINO features without any modifications? These results would help understand how difficult the tasks are and what is the trade-off between using disentanglement vs not using it.\n* It is possible that DINO features are already disentangled and all what the proposed strategy is doing is assigning names to some of the factors of variation that DINO can detect. Therefore, the disentanglement is not really happening in the VAEs but rather obtained from a model pretrained at large scale. What type of experiment can the authors design to test this hypothesis?\n* If the hypothesis above is not rejected, the value of proposed methods is limited to only annotating factors of variation rather than identifying them in a weakly supervised manner to then being transferred." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Referring to the weaknesses noted above, I find the claimed contributions of this paper not sufficiently convincing. Could the authors provide a more compelling explanation of their main contributions, particularly addressing:\n1. Why DRL is specifically suited for microscopy image analysis.\n2. What novel challenges or requirements this domain brings to DRL.\n3. How their approach advances the theoretical or methodological aspects of DRL beyond simple application." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The manuscript is well-written and easy to follow, with clear organization and logical flow.\n2. The application of weakly-supervised DRL to real-world image analysis represents a promising and valuable research direction." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper addresses the interpretability challenge in microscopy image analysis using deep learning approaches. The authors propose a novel methodology based on Disentangled Representation Learning (DRL) to enhance model interpretability while maintaining classification performance. The approach leverages transfer learning from synthetic features and is validated across three diverse microscopy domains: plankton, yeast vacuoles, and human cells. The growing volume of microscopy images due to technological advances has necessitated automated analysis methods, yet interpretability remains crucial for practical applications in fields such as diagnosis and environmental monitoring. The authors demonstrate that their DRL framework successfully balances the trade-off between model accuracy and interpretability in microscopy image classification tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.The scope of this work appears too narrow, focusing solely on microscopy images. The proposed approach might be more convincing if demonstrated on natural images as well.\n2.The authors fail to adequately justify why DRL should be specifically applied to microscopy image analysis. Furthermore, they do not clearly articulate whether this specific application domain poses new challenges or requirements for DRL that could lead to innovative solutions. The authors' insights into these aspects are not well presented.\n3.Given the lack of compelling insights, this work appears to be primarily an application of existing DRL methods without significant methodological or theoretical innovation. This level of contribution may not align with ICLR's focus on novel methodological and theoretical advances in machine learning.\n4.The paper appears to lack comparative experiments. While the disentanglement scores might be novel evaluation metrics, the absence of comparisons for classification performance is particularly concerning and unreasonable." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "Most of my questions are related to major weaknesses.\n\nWhat specific contributions does this paper make beyond applying DRL to microscopy images? It would be helpful if the authors could clarify what is novel in their approach and how it advances the state-of-the-art in microscopy image analysis beyond existing techniques.\n\nWhat are alternative approaches the authors could have used for comparison? \n\nMetric explanations (e.g., OMES, MIG, DCI and balanced accuracy) are mostly missing. Could the authors clarify these metrics, ideally using mathematical notation and provide justification for using them?" }, "rating": { "value": 1 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "The paper explores the application of an existing DRL framework to the specific domain of microscopy images. This idea is interesting as it shows a potential pathway for combining DRL with microscopy image analysis." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a Disentangled Representation Learning (DRL) approach to improve interpretability in microscopy image classification. By pre-training on synthetic data (Texture-dSprite) to capture factors of variation, the authors apply these learned representations to real-world microscopy datasets (Plankton Lensless, Plankton WHOI15, Budding Yeast Vacuoles, and Sipakmed Human Cells). Their method aims to support model interpretability while achieving high classification performance with gradient-boosted trees and MLPs for downstream analysis." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "A significant weakness as it seems, is the absence of a comparison with other similar methods. The paper presents only one framework and does not discuss or evaluate alternative approaches, which weakens the case for this framework’s efficacy or advantage over existing methods.\n\nThe contributions of the paper in terms of novelty are unclear. The study applies an existing DRL approach to a new domain but does not appear to introduce any fundamentally new concepts, techniques, or substantial modifications to existing methods. The only apparent novelty - the application of DRL to microscopy imaging does not suffice. This limits the potential impact and originality of the work.\n\nThe paper’s presentation suffers from numerous issues that impede readability and clarity:\n1. There are instances of informal languages, such as the use of “thanks.”\n2. The text contains multiple errors at the word, sentence, and structural levels, which disrupts the reading experience. Sections like Section 2.2 (“Disentanglement Evaluation”) resemble output generated by ChatGPT and lack rigorous academic polish.\n3. Figures appear low-resolution, with inadequate explanations in captions. Captions should be comprehensive and self-contained, but here, they lack essential details, e.g., explanations of metrics like OMES and balanced accuracy.\n4. The use of multiple highlight types (underscoring, bold, italics) is excessive and distractive. Minimal highlighting would improve readability and make essential points more accessible.\n5. Important metrics are either not explained in the text or lack adequate definitions in the captions, leaving readers uncertain of their meaning. This omission impacts the study’s reproducibility and overall clarity." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "- In Fig, 1, what is exactely fine-tuned and how ?\n\n- How is an RGB image directly fed to the classifier (GBT and MLP)?\n\n- In line 322, the authors state \"We can observe that after finetuning, it may change, nicely adapting to the specificity of the dataset, where scale and texture are more relevant.\", It is unclear for me why scale and texture are more relevent then \"scale and shape\", as it is the case before fine-tuning.\n\n- The proposed evaluation metrics (ex:OMES) are unclear.\n\n- The authors do not compare their method to any other work, having a solid baseline is important.\n\n- The used classifiers (GBT and MLP) are very simple, more sophisticated ones should be used (CNNs based for example).\n\n- Inputting an RGB image to the classifer is unclear as it is well-established that deep features (in this cas the features extracted by DINO) have more important patterns.\n \n- To assess the quality of the representation, the authors realied on classification. While a good representation can lead to a better accuracy. A good representation does not necessary mean a disentangled one.\n\n- Using the accuracy only to measure the quality classification performance is not enough.\n\n- The figures are small and the captions are not clear enough.\n\n- In Figure 6, the OMES indicates that the proposed method does not lead to better disentanglement." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "- The paper addresses a significant challenge in representation learning: disentanglement, which plays a pivotal role in improving the interpretability of classifiers, particularly in the context of biological images." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this paper, the authors propose to use a disentangled representation learning framework to enhence model interpretability for microscopy image classifications. The method is based on fine-tuning a model trained on synthetic images, the proposed framework is tested on some microscopy images datasets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The proposed approach is not well explained. Indeed, the method proposed by the authors learns a disentangled model with weak-supervision using Ada-GVAE on a synthetic dataset and then fine-tune it on\n a microscopy datasets. However, it is unclear why Ada-GVAE is choosed and how is the model fine-tuned.\n\n- The difference between the proposed method and Dapueto et al is unclear.\n\n- The authors claim that the disentanglment learned from a synthetic images can be transferred to microscopy images, such claim should be theoretically and empirically evidenced.\n\n- The paper is not well organized, for instance a \"Related Work\" section should be added. Two different sections (2.2 and 3.5) have the same title \"DISENTANGLEMENT EVALUATION\"." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We transfer a disentangled representation from a Source dataset to real-world Target dataset reaching a compromise between classification accuracy of the downstream task and interpretability in microscopy images" }, "_bibtex": { "value": "@inproceedings{\nanonymous2024disentangled,\ntitle={Disentangled representations of microscopy images},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=0iAZYF9hrl},\nnote={under review}\n}" }, "abstract": { "value": "Microscopy image analysis is fundamental for different applications, from diagnosis to synthetic engineering and environmental monitoring. In the last few years, the number of available images has been constantly growing, thanks to technological advancements, pushing toward the development of automatic image analysis methods based on deep learning. Although deep neural networks have demonstrated great performance in this field, interpretability — an essential requirement for microscopy image analysis — remains an open challenge. \nThis work proposes a Disentangled Representation Learning (DRL) methodology to enhance model interpretability for microscopy image classification. \nExploiting benchmark datasets coming from three different microscopic image domains, including plankton, yeast vacuoles, and human cells, we show how a DRL framework, based on transfer learning from synthetic features, can provide a good trade-off between accuracy and interpretability in this domain." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Microscopy images", "Disentangled representations", "Transfer learning", "Interpretability" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/89f2963d7234a9b02fc373335d78dd67c3e84695.pdf" }, "presentation": null, "primary_area": { "value": "applications to physical sciences (physics, chemistry, biology, etc.)" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/423740d03cb2fefca8e8ba22ac7026fb26e4a65c.zip" }, "title": { "value": "Disentangled representations of microscopy images" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Statistical Estimation Methods for Timestamps: Which specific statistical methods are used for estimating timestamps in the absence of explicit temporal markers?\n\nUnion of Tacoustic and Tvoiced: In combining Tacoustic and Tvoiced, what is the methodology for performing this union? Is Tvoiced identical or related to another variable, such as St?\n\nFormat and Extraction of Visual Details: What is the format of the visual scene details (e.g., Vscene, Vplace, Vtime), and how are these extracted and integrated into the memory graph?\n\nDefining Key Events and Hierarchical Organization: How are key events identified, and what hierarchical structure is used to organize these events?\n\nRelation between Taudio and Tcombined: Is Taudio equivalent to Tcombined, or is there another relationship between these variables?\n\nTask Categories for Text Summarization: How are text summaries grouped into broader task categories (e.g., meetings, lunches)? What criteria and process are used to define these categories?\n\nSimilarity Calculation in Equation 8: Equation 8 is intended to measure similarity between an event and multiple episodes, but it isn’t clear how it accomplishes this. Could you clarify how this calculation works?\n\nLocation Weight Definition: How is \"location weight\" defined, and how does it differ from location similarity in the model?\n\nTemporal Parameter in Equation 14: In Equation 14, should the parameter be (t-k) instead of just t? If not, what purpose does the current form of the equation serve?\n\nMeaning of \"Agent Comprehends\": In line 274, it says the \"agent comprehends\" something. Does this imply processing by a language model, and if so, could you clarify which model is used?\n\nDefinition of the Set Du: How is the set Du defined in the context of the framework?\n\nSimilarity Function in Line 283: Which similarity function is used in line 283, and what factors are considered?\n\nRole of w and l Functions: In line 287, the w and l functions are mentioned. Could you elaborate on their roles within the memory retrieval mechanism?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "I think the paper is trying to address a very important problem of how can an autonomous agent keep memoring new experienes and the recall those flexibly based on the context, question or query. Specifcally I see two main strengths. \nPretrained Models: The model builds on pretrained models that help with extraction of compenents, but this also means the system does not require pre-training on every specific scenario from scratch, which is a significant strength as it allows flexibility across contexts.\nDataset Diversity: The authors evaluated the system on multiple datasets, demonstrating a broad application range, although it should be noted that most of these datasets were developed by the authors." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents Episodic Memory for Cognitive Agents (EMCA), a framework designed to support memory retention and retrieval in cognitive agents. EMCA models episodic memory using a graph-based structure that incrementally stores and organizes multimodal experiences—such as speech, vision, and non-verbal cues—without pre-training on specific scenarios. This approach enables agent to keep adding new experiences continuously from data. This supposedly allows for flexible temporal reasoning across different datasets. EMCA’s dynamic memory graph builds semantic and temporal connections, enabling context-aware retrieval and clustering of memories based on query relevance. The framework aims to improve task execution , and reasoning by recalling contextually significant past events. Empirical tests reported indicate that EMCA adapts to real-world data, demonstrating good recall in unpredictable settings." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Despite tackling an important problem, the paper suffers from serious clarity and coherence issues that obscure its contributions and weaken its scientific rigor. The presentation is fragmented, key concepts are inadequately explained, and essential technical details are missing, all of which make it challenging to assess the model’s validity and potential impact. Specifically, the weaknesses include:\n1. The authors claim EMCA encodes data in a way that resembles human memory, but there is no evidence or detailed explanation to support this claim from a neural encoding way. Such claims should be toned down. You should instead emphasize on the 'what', 'where' 'when' organisation from a psychological perspective of episodic memory. \n\n2. Insufficient Motivation: The introduction section does not adequately establish the necessity of this system or why it improves upon existing learning frameworks for cognitive agents. Additional motivation for the need of an episodic memory for a cognitive agent would help contextualize EMCA's contributions.\n\n\n3. Minimal Related Work Discussion: Essentialy the model is an encode and retrival model with some dynamic reorganistion. The related work section is sparse and lacks comparisons to key formal methods like Hopfield networks or other established models in episodic memory encoding and retrieval. A more rigorous comparison to established human memory models would also strengthen the paper.\n\n4. Unclear Implementation and Integration Details: Although multiple models and methods are mentioned, the paper lacks a cohesive description of how these components integrate within the system. Critical details such as model architecture, parameter settings, and processing pipelines are absent, making it difficult to assess or replicate the work. A system architecture diagram, a table of key parameters, or pseudocode for the main processing pipeline would help.\n\nVague Statistical Estimation Methods: The paper mentions the use of statistical methods for estimating missing timestamps but does not specify which methods were used, leaving an important aspect of the framework unexplained.\n\nSurface-level Comparison with Temporal and Knowledge Graphs: The comparisons with temporal and knowledge graph structures are brief and lack depth, offering limited insight into how EMCA differs from or improves upon these existing approaches.\n\nUndefined Terminology and Variables: Certain terms and variables (e.g.,\"key events\", \"location weight,\" \"subjective temporal timescales\") are introduced without sufficient explanation or definition, reducing clarity.\n\nOverreliance on Custom Datasets: While the use of various datasets to evaluate EMCA is a strength, most of these datasets were developed by the authors, which could indicate potential biases in testing and validation.\n\nLimited Explanation of Retrieval Policy: The retrieval policy and memory clustering mechanisms, while central to EMCA’s functionality, are described only briefly. A more detailed explanation would clarify how these mechanisms adapt to different query types and scenarios." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "- How was the dataset constructed? This is a major contribution that is not discussed.\n- What does a successful or unsuccessful example look like? I would recommend looking at [2] or [3] above to see how to discuss dataset creation in such a setting.\n- In Table 1, what metric is being used? It does not say in the caption or the table. It should be re-iterated in the table itself.\n- Is a forgetting mechanism necessary, or would it be more like a memory aggregation mechanism so that retrieval is still efficient?\n- How does the incremental storage/retrieval scale as the number of episodes change? This result is not displayed in the paper but I would argue is very important. If you only used say 10 episodes of EM instead of 181, is there a difference in performance? This would directly support contribution number 2 in Section 1 of your paper\n- What is \"Time complexity\" in Table 3? Time complexity of BFS is O(V+E), but here a number is used instead. Authors should use \"retrieval time\" or something similar instead of time complexity.\n- Can the authors describe the main takeaways of Section 4? I still do not understand the insights this section is supposed to provide.\n- Also, there are some questions/concerns in the weaknesses section\n\n\nOverall, I like the paper. But I think there are a lot of issues with how to content is shown to the reader that makes the paper's contributions fall flat. In its current state, I would recommend rejecting the paper, but if the authors address my concerns above, I believe I would lean more towards accept." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The Big Bang Theory and the Agent datasets seem very useful! \n- Their results indicate that their method performs better than other graph-based approaches. \n- This area is a growing field, especially as robots and agents become more capable and need better ways to scale their context. And their graph-based approach seems to be a meaningful contribution\n- Table 2 is interesting, as it showcases that some of their questions require access to vision, acoustics, and/or dialogues. This result would be better if we knew what the Big Bang Theory and Agent dataset contained." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors introduce Episodic Memory for Cognitive Agents (EMCA) that models episodic memory based on a graph-structure. This allows them to incrementally store memories and retrieve experience. They also can cluster memories, have dynamic retrieval, and can handle temporal reasoning. Memory is structured as a graph, where each episode contains characters, temporal elements, location, and events. The edges of this graph can be temporal or semantic. They then build a retrieval system that can handle contextual, temporal, and spatial queries." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "## Main weaknesses\n- After reading the full paper, I am not sure what the actual task is. What commands does the master ask? I understand the approach, but not what task it is specifically solving. Is the output the location of a specific memory? Or is it a free-form text answer?\n- In terms of the structure of the paper, I do not think that the text space used for discussing how signals are captured (section 3.1) is very useful. I would argue this is the case with a chunk of this paper's equations, where it simply adds space when it does not need to, and it makes the paper more difficult to read. I would recommend condensing the information and matching ICLR's 9-page recommendation as opposed to 10. Similarly, there is an excessive use of new lines at arbitrary positions.\n- There is no discussion on how the Big Bang Theory and Agent datasets were constructed. This on its own seems like a major contribution on its own. I would recommend the authors remove much of the superfluous equations and newlines (and move that into the appendix), and put more of an emphasis on this dataset component.\n- I think I like what Section 4.1 is implying about combining a \"master's\" memory with that of an agent's, but it is not presented very clearly, and I do not see a connection with temporal graphs like the subsection title suggests\n- Nor do I see how this section is a \"theoretical comparison\"\n- \"Master\" terminology in line 346 is confusing, and should be introduced earlier in section 4.1. Also, the term \"master\" is generally frowned upon in these settings, so I would recommend a different term\n- Results are poorly presented\n\n\n## Formatting/Clarity issues:\n- Check for spaces after periods or colons throughout the paper.\n- Line spacing is odd in much of the paper\n- Figure placement should ideally be on the top or bottom of a page, not in the middle with paper text above and below the paper. This makes the paper difficult to follow\n- Figure 4, the legend has oddly shaped circles\n- Figure 5's result is good, but it should not be a line graph with x axis being method and lines being the dataset. Instead it should be datasets on the x-axis and methods on the y-axis\n- The results in section 5.1.1 are all discussed in a single paragraph. Is that all the main results? I would recommend splitting this up into a few bolded mini-sections and showing the main takeaway of each figure along with highlights on how the method performed, possibly with qualitative results.\n- The focus of the introduction on falls a bit flat. Rather than focusing on the historical definition of episodic memory, focus more on how people have been engineering and building these kinds of systems. Focus on why other systems do not work, and why yours does.\n- I would recommend larger fonts for the figures; they are difficult to read in the paper.\n## Citations\nOther relevant concurrent work on memory in robotics, some of which use graphs while others do not.\n[1] Xie, Quanting, et al. \"Embodied-RAG: General Non-parametric Embodied Memory for Retrieval and Generation.\" arXiv preprint arXiv:2409.18313 (2024).\n[2] Anwar, Abrar, et al. \"ReMEmbR: Building and Reasoning Over Long-Horizon Spatio-Temporal Memory for Robot Navigation.\" arXiv preprint arXiv:2409.13682 (2024).\n[3] Bärmann, Leonard, et al. \"Episodic Memory Verbalization using Hierarchical Representations of Life-Long Robot Experience.\" arXiv preprint arXiv:2409.17702 (2024)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. Will the code and datasets be made publicly available?\n2. How does the system maintain temporal consistency without explicit time markers? Can you provide quantitative results comparing temporal reasoning accuracy with and without explicit timestamps?\n3. What are the specific model architectures and hyperparameters used?\n4. What are the computational costs for memory retrieval at different scales?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "originality: graph-based episodic memory structure multimodal processing without pre-training, dynamic clustering\n\nquality: comprehensive testing on multiple datasets, benchmarking against exsiting methods, systematic component analysis in ablation study\n\nclarity: clear problem formulation, good visual aids explaining complex concepts\n\nsignificance: adresses crucial challenges in cognitive AI, social robot applications, potential impact on memory assistance systems\n\nkey innovations: \n1) removes pre-training requirements\n2) enables continuous learning\n3) provides real-time processing campability" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents Episodic Memory for Cognitive Agents (EMCA), a novel framework that enables AI systems to retain and utilize past experiences through a graph-based memory structure. The key innovation lies in its ability to: 1) Process multimodal data (vision, speech, non-verbal cues) without requiring pre-training; 2) Dynamically build and update memory representations through a graph structure with semantic and temporal connections; 3) Adapt to complex environments through continuous learning from interactions." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Scalability limitations: Does the graph structure grow exponentially with experiences? What are the computational costs?\n2. More details about the Big Bang Theory dataset are needed.\n3. More implementation details about the method in section 5.0.2 should be provided.\n4. Forgetting is one of the key problems in memory systems - how does the paper assess and handle memory 5. retention and decay?\n5. What is the real-time performance?\n6. The paper tries to claim episodic memory for agents and robots, but robotic interaction with the environment is different from and much harder than agent interaction in a virtual environment. It is better to make a clear definition and scope. For instance, in L391, \"robot's episodic memory\" should be \"agent's episodic memory\".\n7. Can authors provide the code and dataset for evaluation?\n8. Repeated paragraph: L054-L062\n9. Subtitle formatting issue in line 480" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "When this method gets deployed, there would potentially be ethics concerns - as pointed out by the authors.\nIn the submission, it just uses datasets (established and generated form a TV series) rather than real user data, so no concerns." }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "- The paper seems to miss all details on how the responses are generated, but the whole experiments are based on evaluating the answers. Please provide a detailed explanation of your response generation process, as this is crucial for understanding and evaluating the experimental results.\n- Sect. 3.2.4: This seems to result in just 'jumping' nodes. Conserving the chronological structure is a nice property, but the real question/challenge rather is on the side of the module that determines if there is any harm in 'skipping a step', e.g. when thinking about some of the uses-cases 'predict ... next activity' you mention\n- Quite a lot of unclear details\n - Eq (1) vs (2): the difference between S_T and T_voiced is unclear\n - How is T_acoustic generated\n - Sect. 3.1.2, 3.1.3, and 3.2, l. 218: at places the paper sounds like everything is represented as 'text', at others it seems to be a mixture of text and other embeddings. It would be great to explain earlier on what it stored as what\n - l. 184: \"organized by place, characters, and events\" raises the question where those come from - that is explained later in the text\n - l. 220: \"relevance of these tasks is assessed\" - again HOW?\n - l. 212: what are the implications/limitations resulting form using a simple metric like cosine similarity?\n - Eq (17) sim seems undefined\n- Fig. 2 and Sect. 4.1: I didn't get the terminology \"master\". Maybe that term can be avoided altogether (similar to https://learn.microsoft.com/en-us/style-guide/a-z-word-list-term-collections/m/master-slave )\n- The paper comes across as unpolished\n - missing and extra spaces - e.g. in the title and abstract\n - the template uses natbib, not using correct commands for references \\citep \\citet (and resulting repeats of names) makes it painful to read\n - LaTeX has different opening and closing quotation marks\n - paragraph above \"Contributions\" is double\n - broken sentences in Sect. 5.1.2\n - a few references with incomplete info (e.g. l. 300 \"as shown in 4\", l. 191, l.468)" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper tackles and interesting problem. The architecture is described in an intuitive way and seems sound and novel. The experiments are reasonably extensive with comparisons to baselines, multiple datasets, and ablations and show very promising results." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a novel method for storing multi-modal episodic memories. The design takes inspiration from a model of human memory. More specifically, the paper introduces a novel type of memory graph with different types of connections. The method compares favorably against baselines in experiments." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- I found the paper to have some \"false advertising\":\n - Some of the introduction, motivation, and discussion circles around 'robots', I didn't see anything specific to robots in this paper. Yes, it could be integrated into a robot, but the method would equally well work for a body cam, social agent, etc. In robotics there is quite an extensive literature on 'lifelong learning' that covers some of the same challenges: what to memorize, how to store and to retrieve, how to generalize, what to forget, etc.\n - The title says 'learning'. At least according to my definition there is no learning in the paper. It proposes a way to store information and to retrieve it, so that would correspond to 'memorization' (=rote learning), while learning implies understanding the information and being able to apply it to new situations. The method could serve as a starting point for learning, but in the current paper that doesn't seem to be present in the method nor in the experiments.\n - The introduction makes it sound like a general method. But I have a few doubts about that. There are quite a lot of design choices, and some of choices in Sect. 3 seem rather specific. In the end the method is evaluated with a question answering task. What would need to change for a different task, say for a robot learning low-level movement control? Please clarify the generalizability of your method. Specifically, it would be nice if you could discuss how your approach might be adapted for different tasks (such as robot movement control), or explain any limitations in its applicability to other domains.\n- Section 3 describes the method. It remains however very much on the level of HOW, rather than providing many insights in the WHY (reasoning behind design choices, consequences of design choices) and it isn't always clear what is a core part of the method and what is an implementation detail. Please provide more explanation for the key design decisions, discussing the rationale behind these choices and their potential implications. Additionally, please clearly distinguish between core methodological components and implementation details.\n- Fig. 5 isn't very convincing (except for Arigraph)\n- Memory requirement would be another interesting metric for comparing methods.\n- The paper mentions the missing forgetting mechanism as a major limitation. Related to that it also leaves the question on 'what to store' unanswered. It reads like everything is stored, even if it is effectively a duplicate. I believe the real challenging question that needs to be solved is the memory management: what to store, what to consolidate/merge, what to forget, etc. Without those a memory representation is of limited value, and it remains unclear to me how suitable the proposed architecture is for extending it in that way - or if we would be better off redesigning it from scratch.\n- The method relies on various models to extract features and to summarize things before storing them. I don't believe there is a 'one size fits all approach' but how to best do that depends on the retrieval task and the type of downstream tasks you have. The paper does not provide any indications on how to deal with that.\n- There are a whole lot of design choices in this paper, an ablation on only the modalities and search methods seems a bit limited. There also is no sensitivity analysis for e.g. the clustering thresholds\n- This seems to be a rather complex system, which makes reproducing results very difficult, with probably quite lot of additional implementation and setting details. I couldn't find any promises to release code (or at least a detailed appendix), which would have alleviated this concern." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024learning,\ntitle={Learning through experience:Episodic memory representation for cognitive agents},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=0iXfS9Smqf},\nnote={under review}\n}" }, "abstract": { "value": "As the demand for intelligent robots and cognitive agents rises, the ability to retain and utilize past experiences through episodic memory has become crucial, especially for social companion robots that rely on previous interactions for task execution. To address this, we introduce Episodic Memory for Cognitive Agents (EMCA), a novel framework that advances knowledge representation by integrating real-world interactions. EMCA enables agents to adapt to complex environments by learning from tasks, interacting with humans, and processing multimodal data—such as speech, vision, and non-verbal cues—without pre-training on specific scenarios.\nEMCA models episodic memory through a graph-based structure , allowing for incremental storage and retrieval of experiences. Each interaction or event enriches the memory graph, supporting continuous learning and adaptation without extensive retraining. This human-like memory formation optimizes the agent’s ability to retrieve relevant information for tasks like localization, planning, and reasoning based on prior experiences.Unlike conventional models relying on temporal markers or recurrent patterns, EMCA encodes data like human memory, allowing reasoning across diverse scenarios regardless of temporal patterns. The framework dynamically builds a memory graph with semantic and temporal connections based on the agent’s experiences, promoting flexible temporal reasoning. It also introduces mechanisms for clustering new memories and a dynamic retrieval policy that adjusts based on context or query type, ensuring robustness even in unpredictable scenarios. Empirical tests show EMCA adapts effectively to real-world data, offering reliability and flexibility in dynamic environments." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Episodic Memory", "Bio inspired Robot learning", "incremental Memory structures" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/dc76713003b768e3eb75d5d5ff0cf6fbaeb978ef.pdf" }, "presentation": null, "primary_area": { "value": "transfer learning, meta learning, and lifelong learning" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Learning through experience:Episodic memory representation for cognitive agents" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Visually, it seems difficult to reconcile the large difference between, say, Figure 2 top left panel and Figure 1 top left panel. It seems that on the TUP dataset, high risk is close to maximizing treatment effect, the highest values being around percentile 80. However, Figure 2 shows utility 15000 vs 5000 for treatment effect based vs risk vased. Can you explain?\n\nSimilarly, treatment effects for NSW seem to be largely constant around 1000, but Figure 2, second row left panel shows a massive advantage for treatment effect based targeting. This seems to be about a factor 7 (1.75 vs 0.25). Where do these large effect differences come from given that the treatment effect curve is essentially constant? None of the treatment effects seem to differ by more than a factor 2.\n\nI tried to figure this out by looking through the code, but I couldn't find code that generated Figure 2. So, I don't actually know where it came from and what it shows. Maybe I missed it. On that note, it would be very helpful to clean up and document the code for the final version. As is, it's very hard to follow." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "(1)\n\nThe comparison of these two targeting strategies is an important problem. I'm glad to see that the authors study this question. Despite its significance, there hasn't been much reliable insight so far. I'd love to see more work in this direction! I'm weighing this strongly in my evaluation.\n\n(2)\n\nThe paper is very clearly written. The authors narrate a compelling story about the advantages of targeting based on treatment effect estimates, even if they are biased." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors compare the utility of targeting interventions based on estimated treatment effects with the utility of targeting based on predicted risk. The former is the method of choice from a utilitarian perspective. But practitioners and policy makers often choose the latter due to its simplicity and in cases where the normative goal is to assist those in greatest need." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "(1)\n\nAs compelling as the story is, I find the results less than conclusive. Looking at Figure 1, it looks like the confidence intervals around treatment effects are generally strongly overlapping across the entire x-axis of baseline risk. The one exception is the STAR dataset, where the lowest and highest point estimate have barely non-overlapping confidence intervals. As a result, another story consistent with the data is that effects are generally not so heterogeneous as conventional wisdom from the econometrics literature has it.\n\n\n(2)\n\nThe results in Figure 1 are actually a fair bit more favorable towards risk-based targeting than the introduction of the paper had me believed. Treatment effects generally increase with risk. Targeting the 80th to 90th percentile of risk generally seems to capture high treatment effects across all datasets. So, another story could be that we should exclude the most extreme values of risk from targeting, but other than that risk-based targeting sort of works.\n\n\n(3)\n\nI found it rather confusing to have unnormalized utility values on the y-axis. After all the sample size is rather arbitrary and does not correspond to the population-level utility obtained if the policy maker were to implement the given approach. Along the same lines, I found it difficult to reconcile Figure 1 and Figure 2. See question below.\n\n\n(4)\n\nIt would've been great to include datasets with real world confounding rather than the simulated confounding. My experience is that existing CATE estimation methods don't do very well in non-RCT settings. Might there be an advantage to risk-based targeting in non-RCT data?\n\n\nSuggestion:\n\nIt seems to me that the story is much less certain than the abstract and introduction make it sound. I would therefore appreciate it if you could indicate a greater level of epistemic uncertainty in the writing throughout. I don't think this would hurt the paper at all. As is, though, I'd worry that your writing suggests the question is essentially closed conclusively which would actually discourage additional work in this direction." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Modeling Real-World Confounding: Could the authors expand on how their confounding\napproach aligns with real-world biases encountered in observational data? A discussion on potential limitations in modeling confounding factors might aid readers in interpreting results for specific applications.\n\nEthical Implications: How might the authors' conclusions address ethical concerns,\nparticularly in terms of balancing fairness (people at the most rish) with effectiveness\nwhen treatment effect targeting benefits some groups more than others? Because the paper concludes that a treatment-effect-based method is preferable to a risk-based method, it raises\nthe natural question of whether individuals at the highest risk should be prioritized, or if those with the greatest\npotential treatment outcome should be targeted instead.\n\nData Quality: Could the authors provide more information on the reliability and source of the\nAcupuncture Dataset?\n\nConfidence Interval Estimation: In Equation (5), a biased estimator is used. Could the authors\njustify why this choice was made? Was it primarily for computational efficiency, or are there\nother reasons for using a biased estimator in this context?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "Scope: The work’s target is evident in its cross-domain empirical evaluation, making it applicable to multiple policy\nareas. The paper introduces innovative use of biased treatment effect estimates to assess targeting efficacy.\nQuality: Methodologically, the paper is solid, employing credible RCT data and a robust approach to measuring\ntreatment effect heterogeneity. The use of doubly robust estimation and varying social welfare functions adds\ndepth to the analysis." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper explores targeting strategies for allocating limited resources to maximize social welfare in areas like\nhuman services, healthcare, and education. Specifically, it compares risk-based targeting, which prioritizes\nhigh-risk individuals, with treatment-effect-based targeting, which uses machine learning models to estimate who\nmight benefit the most from interventions. Using five real-world randomized controlled trials (RCTs) across diverse\ndomains, the paper concludes that even biased estimates of treatment effects generally outperform risk-based\ntargeting. This finding suggests in addition to the widespread reliance on risk-based approaches, policymakers\ncould incorporate treatment effect estimation when feasible." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Grammatical errors: there are notable grammatical errors in spelling and following the formatting of ICLR\ninstructions.\n\nScope of Treatment Effect Estimates: The reliance on simulated confounding could be more thoroughly justified;\nreal-world application requires consideration of domain-specific biases that could vary across different policy areas\nPotential Ethical Implications: Since the work could influence resource allocation in sensitive domains, further\ndiscussion on ethical considerations regarding inequality and bias would strengthen its impact and application." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Feel free to further discuss the first two weaknesses mentioned above.\n\nI have seen papers, such as [1], that discuss the complexities of estimating allocation welfare under budget constraints. The issue is that these constraints introduce dependencies between individual estimations, complicating the evaluation process. How is this relevant to your setting?\n\nShouldn't the results also depend on how many individuals are treated in each scenario? My intuition (supported by studies like [2]) suggests that the relative budget plays a significant role in determining which type of allocation mechanism is optimal.\n\nSince you studied a wide range of settings, did you find any recommendations that vary across these contexts? For instance, are there conditions under which risk-based targeting performs better?\n\nCould you elaborate on the choice of outcome of interest in the TUP dataset? It seems there’s an implicit assumption that individuals who show a larger increase in expenditure are more deserving of intervention. Why should this be the case? In contrast, a risk-based approach might assume that those who would have lower expenditures in the absence of interventions are more deserving. Could this actually be a more appropriate assumption?\n\n[1] Improved Policy Evaluation for Randomized Trials of Algorithmic Resource Allocation\n[2] Allocation Requires Prediction Only if Inequality Is Low" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "This is an important issue, and I’ve personally never found a clear guideline on how best to approach it. I found the study original and of good quality.\n\nThe paper does a commendable job of incorporating datasets from diverse domains. \n\nIt also simulates the impact of unobserved confounders and considers the possibility of policymaker bias toward risk-based targeting." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper compares ''risk-based targeting'' and ''treatment effect-based targeting'' across five datasets from RCTs in different domains. In risk-based targeting, the planner targets individuals based on their baseline risk, which is the expected outcome in the absence of treatment. This approach, commonly used by practitioners, does not account for causal effects. In contrast, treatment effect-based targeting involves first estimating the Conditional Average Treatment Effect (CATE) and then prioritizing individuals accordingly, often with the help of machine learning. A key concern is that these methods may introduce bias in the presence of unobserved confounding. The paper aims to provide empirically grounded guidance for navigating this tradeoff.\n\nFirst, the paper investigates the extent to which baseline risk serves as an effective proxy for targeting. To do this, treatment effects are estimated at different levels of baseline risk, revealing a surprisingly complex and not necessarily monotonic relationship between treatment effect and baseline risk. Second, the paper examines the potential additional gains from targeting based on treatment effect. Both utilitarian and Nash welfare are used to compare these mechanisms, showing that targeting based on estimated treatment effect can be up to three times more effective in nearly all cases. This remains true even when confounding is introduced into the CATE estimation and when the policymaker's preference for risk-based targeting is encoded in the social welfare function." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "My main concern is the reliability of comparing the estimated welfare from risk-based versus treatment effect-based targeting. The issue with the latter approach, particularly in resource-scarce settings, is that it first estimates CATE, then selects individuals with the highest estimated CATE given the available budget, and finally estimates the welfare of treating these individuals using the same estimated CATE. Even if the CATE is unbiased, this procedure can lead to severely biased welfare estimates in resource-constrained settings.\n\nIn Figure 1, the paper provides treatment effect estimates for each baseline risk percentile, along with confidence intervals. What exactly are we looking for here? If the goal is to rule out a monotonic relationship between baseline risk and treatment effect, none of the figures offer statistically significant evidence to support this.\n\nI also think the results should intuitively depend on the budget but this is not discussed in the paper. For instance, if the budget is very small, treatment effect-based allocation can be even more effective in exploiting heterogeneity. \n\nOverall, the writing is strong. I noticed the following minor typos: page 2 (potentially-based -> potentially-biased), page 4 (policy -> policy), page 5, second equation, page 7, Equation 4.5 is referenced which does not exist" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 4 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "Fig. 2 is great, almost telling the story of the paper without the audience needing to read anything else. What would make it even better is replacing \"Percentage of data removed\" with something like \"Higher values mean more confounding in CATE estimates\". Currently readers have to read the methodology to understand the x axis, but a better label would mean they could understand the x axis without knowing exactly how the confounding was introduced, and read the methodology if they wanted more details." }, "rating": { "value": 10 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "This is a simple, well-executed paper making a very important point. As the authors note, most algorithmic decision making approaches are based on risk or some manipulation of risk in pursuit of fairness goals. This paper demonstrates the problem with that approach across a wide range of interventions.\n\nThe authors anticipate the key criticisms of the policy based approach (that RCT data is hard to find and so CATEs may be biased, and that equity preferences might provide a non-utilitarian reason to prefer risk-based targeting), and provide compelling evidence that even under significant confounding or equity preferences one should still prefer allocations based on CATEs." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper studies a common risk-based approach to treatment allocation applied to a diverse range of treatments with randomized controlled trial data available. The authors find that allocating treatment according to risk (ie, bad outcomes under the no-treat baseline) produces worse outcomes than allocating according to conditional average treatment effect estimates, even when CATEs are biased or the decision maker has a preference for treating high-risk individuals." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The weakest part of this paper is probably the discussion of Fig 1. The authors cite a \"unique trend for each dataset\" (in CATE conditional on risk score) as evidence that the risk-based policy is flawed. Visually, to me, the trends look pretty similar: they look like there is basically no relationship between CATE and risk for most RCTs studied. This is still evidence for the risk-based policy being flawed (so it doesn't invalidate the claim being made), but I think it's a more accurate way to describe the results. I think Fig 2 is more compelling than Fig 1 anyway, so it probably make sense to lead with that result for the most impactful paper." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024comparing,\ntitle={Comparing Targeting Strategies for Maximizing Social Welfare with Limited Resources},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=0iscEAo2xB},\nnote={under review}\n}" }, "abstract": { "value": "Machine learning is increasingly used to select which individuals receive limited resource interventions in domains such as human services, education, development, and more. However, it is often not apparent what the right quantity is for models to predict. In particular, policymakers rarely have access to data from a randomized controlled trial (RCT) that would enable accurate estimates of treatment effects – which individuals would benefit more from the intervention. Observational data is more likely to be available, creating a substantial risk of bias in treatment effect estimates. Practitioners instead commonly use a technique termed “risk-based targeting” where the model is just used to predict each individual’s status quo outcome (an easier, non-causal task). Those with higher predicted risk are offered treatment. There is currently almost no empirical evidence to inform which choices lead to the most effect machine learning-informed targeting strategies in social domains. In this work, we use data from 5 real-world RCTs in a variety of domains to empirically assess such choices. We find that risk-based targeting is almost always inferior to targeting based on even biased estimates of treatment effects. Moreover, these results hold even when the policymaker has strong normative preferences for assisting higher-risk individuals. Our results imply that, despite the widespread use of risk prediction models in applied settings, practitioners may be better off incorporating even weak evidence about heterogeneous causal effects to inform targeting." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "social welfare", "causality", "treatment", "treatment effect", "targeting", "risk", "policymaking" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/dfa365b0c13e73fc971d6779370139dadc5f42f3.pdf" }, "presentation": null, "primary_area": { "value": "alignment, fairness, safety, privacy, and societal considerations" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/b27328a21ff613c511371761b4e0a16628cb18cd.zip" }, "title": { "value": "Comparing Targeting Strategies for Maximizing Social Welfare with Limited Resources" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 2 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "- How does the level of overparameterization affect the EoS phenomenon? This might be an easy extension of Figure 5, by checking if increasing the level of overparameterization, i.e., the ratio $\\frac{d}{n}$, changes the EoS phenomenon. Figure 5 contains $\\frac{d}{n} = 2, 1$, but another experiment on a larger value of $\\frac{d}{n}$ might show if large overparameterization helps in EoS. Note that this is not strictly required but might be interesting." }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "- **Novel insights**: There are several novel and non-trivial insights -- i) quadratic losses can lead to EoS, ii) the diagonal NN model fulfils this requirement, iii) EoS can have regimes with oscillations with increasing magnitude but still converge later, iv) overparameterization might be necessary for EoS on quadratic losses. \n\n- **EoS on diagonal neural networks**: Diagonal neural networks serve as a simple to analyze but still expressive model. Using these models has 3 important advantages -- i) as these are real networks, their claimed phenomenon occurs not just on some theoretically well-crafted model, ii) diagonal NNs remain a good testbed for theoretical analysis of complicated deep learning phenomena, iii) they have provided a proof for EoS on diagonal neural networks, which was missing from existing works (Even et al 2023), and can now be used to verify claimed empirical phenomenon (Even et al 2023). Note that the proof of EoS on diagonal neural networks is highly non-trivial especially for the $\\mu\\eta > 1$ case.\n\n\n- **Presentation**: The paper is easy to read inspite of the heavy notation. The key insights are clearly explained and the figures are very helpful in understanding them. Figure 6, in particular, is a good example to intuitively explain the proof sketch.\n\n\n\n\n**References**--\n- (Even et al 2023) (S)GD over Diagonal Linear Networks:\nImplicit Bias, Large Stepsizes and Edge of Stability. NeurIPS." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors analyze the Edge-of-Stability (EoS) phenomenon for gradient descent on linear regression with the loss function $\\ell(\\langle x, \\beta\\rangle - y)$, where $(x, y)$ is the datapoint with $x\\in \\mathbb{R}^d$ and $y\\in \\mathbb{R}$. EoS phenomenon is observed when GD is run with a stepsize $\\eta > \\frac{2}{L}$ where $L$ is the smoothness constant of the loss. In the EoS regime, GD oscillates rapidly but still converges to the minima, under certain conditions.\n\nExisting works (Ma et al 2022, Ahn et al 2022, Song & Yun 2023) show that for sub-quadratic $\\ell$, GD can enter the EoS regime. This paper shows that for a particular quadratic parameterization, namely diagonal neural networks, GD can enter the EoS regime even for quadratic $\\ell$. The parameterization is $\\beta = \\beta_{w} = w_{+}^2 - w_{-}^2$ and $w= [w_{+}^\\top, w_{-}^\\top]^\\top$, with gradient updates on $w$.\n\n\nFrom Claim 1, for quadratic $\\ell(s) = s^2/4$, they obtain EoS for GD under a single-sample regime for $d\\geq 2, y\\neq 0$ and $x$ non-degenerate. Further, for $d=2, x= (1,x')$ and sparse realizable model $\\beta^\\star = (\\mu, 0), y= \\mu$, with initialization scale $\\alpha$, they obtain two separate regimes even for EoS. \n\nIn the first regime, from Theorem 1, for $\\mu \\eta <1$ and constant $\\alpha$, GD results in both the traditional Gradient Flow regime (GF), without oscillations, and the EoS regime. Here, EoS occurs with damped oscillations. Further, the final solution of GD has generalization error dependent on initialization $\\alpha$.\n\nIn the second regime, from Theorem 2, $\\eta \\mu \\in (1,2)$, with initialization, $x'$, and generalization error dependent on $\\eta\\mu$, GD in EoS might initially have diverging amplitude of oscillations, however, it eventually dies down and after a point converges at a linear rate.\n\n\nEmpirically, the authors show that their model requires overparameterization, as for $d=n$, EoS doesn't occur but for $d>n$, it does. For their case of single sample $n=1$, justifying their choice of $d=2$. \n\n\n**References**--\n- (Ma et al 2022) Beyond the Quadratic Approximation: the Multiscale Structure of Neural Network Loss Landscapes. Arxiv.\n- (Ahn et al 2023) Learning threshold neurons via the “edge of stability”. NeurIPS.\n- (Song & Yun 2023) Trajectory Alignment: Understanding the Edge of\nStability Phenomenon via Bifurcation Theory. NeurIPS." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- **Is subquadratic growth \"necessary\" or \"sufficient\" for observing EoS**? It might be beneficial to state the exact reference for the subquadratic condition. Note that the subquadratic condition was introduced in (Ma et al 2022), which the authors have not mentioned. Further, I'm not sure if (Ahn et al 2023) actually state that subquadratic growth is \"necessary\" for EoS. Assumption A2 and A3 in (Ahn et al 2023), show that subquadratic growth is sufficient for EoS. Similarly, the results in Section 4 in (Ma et al 2022), and assumptions 2.4 and 4.2 in (Song & Yun, 2023) are sufficient for EoS. If I'm missing some details and subquadratic condition is indeed necessary for EoS, can the authors should probably specify the exact theorem, assumption or an argument for this?\n\n- **How large is $\\mathfrak{t} - t_0$** ? In Theorem 2, there are $3$ phases for GD. In the first phase, it has not started oscillations, which lasts until $t_0$. From Lemma 7, $t_0 \\geq \\Omega_{\\mu, \\eta}(\\log(1/\\alpha^2))$, but only for $\\mu\\eta \\in (0,2)$, which includes $\\mu \\in (1, \\frac{3\\sqrt{2} - 2}{2})$. In the second phase, which lasts from $t_0$ to $\\mathfrak{t}$, the oscillations finally start decreasing in magnitude. Lemmas 14 and 15 establish that there exist such a $\\mathfrak{t}$, but not how large it is. As after $\\mathfrak{t}$, we see linear convergence, how long do we need to wait for it becomes an important question. If the authors cannot establish it theoretically, they might argue empirically that $\\mathfrak{t}$ not very large.\n\n- Typo in Line 1537: $|\\lambda_{1,2}| < 1$.\n\n**References** --\n- (Ma et al 2022) Beyond the Quadratic Approximation: the Multiscale Structure of Neural Network Loss Landscapes. Arxiv.\n- (Ahn et al 2023) Learning threshold neurons via the “edge of stability”. NeurIPS.\n- (Song & Yun 2023) Trajectory Alignment: Understanding the Edge of\nStability Phenomenon via Bifurcation Theory. NeurIPS." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. For convergence in the regime $\\mu\\eta<1$, the current analysis assumes period-2 flip sign in residual, and the result essentially states ${\\ell(w_{2t})}_{t\\geq 0}$ converges. I am wondering if the analysis can be extended to deal with more general oscillations without specific periods, as they are closer to the EoS phenomenon in practice [1]. \n\n2. The paper assumes a specific scaling initialization (lines 218 and 238) for subsequent analysis and the authors claim this initialization is to align with the literature on diagonal linear networks. Regarding this initialization, I have a few questions: \n(a) Why in the theory $w_+$ and $w_-$ are of the same scale while in experiment $w_+$ has larger scale than $w_-$? (b) How important is this initialization for ensuring convergence and EoS? (c) How easy is the analysis to be extended to more general or random (e.g. Gaussian) initializations? Just for quick comparison, some existing works can handle general initializations as long as the initial weights satisfy certain conditions [2,3] while they mainly focus on the 1-d case. \n\n3. The result in the $\\eta\\mu>1$ regime (Figure 4, left) suggests that smaller step size allows smaller generalization errors. This seems to contradict a common belief that a large learning rate is beneficial to generalization as it forces the minimizer to be in a flat region. Could authors provide some insights about this difference?\n\n[1] Cohen, Jeremy M., Simran Kaur, Yuanzhi Li, J. Zico Kolter, and Ameet Talwalkar. \"Gradient descent on neural networks typically occurs at the edge of stability.\" arXiv preprint arXiv:2103.00065 (2021).\n\n[2] Ahn, Kwangjun, Sébastien Bubeck, Sinho Chewi, Yin Tat Lee, Felipe Suarez, and Yi Zhang. \"Learning threshold neurons via edge of stability.\" Advances in Neural Information Processing Systems (2023).\n\n[3] Wang, Yuqing, Zhenghao Xu, Tuo Zhao, and Molei Tao. \"Good regularity creates large learning rate implicit biases: edge of stability, balancing, and catapult.\" arXiv preprint arXiv:2310.17087 (2023)." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The paper is overall well-written and easy to follow. It introduces the settings clearly and provides sufficient illustrations to help present results in different conditions/regimes. The main findings are organized clearly. The proof overview is intuitive for grasping the essence of the proof technique. \n\n2. The paper studies implicit bias in the EoS regime for diagonal linear networks in multi-dimensions, which is a significant step forward as most of the prior works on EoS only deal with minimalist examples where the data is assumed to be 1-d, and it is interesting to see the empirical occurrence of EoS depends on the data dimension (non-degeneracy). Moreover, the diagonal linear network setting is closely related to the GD implicit bias literature, so it can potentially connected to wider works." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper studies the implicit bias of GD on quadratic loss and linear models with diagonal linear network parameterization under the EoS regime. The paper first empirically shows that when data is in multi-dimension ($d\\geq 2$), EoS can occur for quadratic loss, while prior works on EoS suggest that subquadratic loss is necessary for EoS to happen. The experiments show that different choices of step size lead to different oscillation types, from GF regime to EoS regime to chaos to divergence. The paper then theoretically studies the parameter convergence (directly related to generalization) in a sparse solution 2-D diagonal linear network setting and provides non-asymptotic rates in various settings. The results show that in the EoS regime when the step size is not too large ($\\eta\\mu<1$), smaller initialization ($\\alpha$) yields a better generalization, while when step size is large ($\\eta\\mu>1$) there will be an error floor that cannot vanish as $\\alpha\\to 0$." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "My major concern is regarding the \"EoS regime\". In the paper, there is no rigorous definition of the EoS regime. Instead, a flip sign condition $r_tr_{t+1}<0$ is used throughout the theoretical statements. However, I doubt whether the sign flips in residual indeed correspond to the commonly referred EoS phenomenon ($\\eta L_t>2$ and oscillations in loss happen along the trajectory). For example, if we minimize $f(x)=\\frac{1}{2}x^2$ with GD, the sharpness is $1$ and the residual $r_t=x_t$ is flipping its sign if we choose step size $\\eta\\in(1,2)$, but the loss is actually monotonically decreasing and the sharpness is always below $2/\\eta$. \nIn the proof overview (Section 5), it is discussed that $|r_{t+2}|\\leq (1-\\alpha)^2|r_t|$ is possible, but there is no statement comparing $|r_{t+1}|$ and $|r_t|$, and there is no statement on the relationship between step size $\\eta$ and sharpness $L$, so we are not sure if EoS is happening or not. In my understanding, it might be the case that only the $\\eta\\mu>1$ case corresponds to the EoS that people refer to. \n\nMinor typos (not exactly weaknesses): \n\n1. In Figure 4 the x-axis has no label, which I guess is $\\alpha$." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "The authors argued that the model considered in this paper is the depth 2 diagonal linear network. Is that possible to extend the theoretical analysis to the model with higher depths?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "In this paper, the authors consider running GD with a large constant step-size to find linear interpolators that admit quadratic parameterization for the one-sample linear regression task. They theoretically proved the on-sample case and extend the one-sample results by empirically finding conditions in the more general n-sample cases. Theses theoretical analysis and empirical results are presented with clear structures." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this paper, the authors consider the task of finding interpolators for the linear regression with quadratic parameterization and study the convergence of constant step-size GD under the large step-size regime. They focus on the non-trivial question of whether a quadratic loss can trigger the Edge of Stability phenomena. The authors show through both empirical and theoretical aspects that, when some certain condition is satisfied, EoS indeed occurs given quadratic loss." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The major weakness of this submission is that the authors only provide the theoretical analysis and mathematical proof for the one sample case. The empirical results from the numerical experiments in section 3 shows the convergence of GD under the EoS regime when the loss function is quadratic. The authors only present the theorems characterizing the convergence of GD when the model considered has dimension $d = 2$. The authors should extend these theoretical analysis to the high dimension cases. The one sample analysis and teo dimensional proof is not enough to explain the empirical results from the numerical experiments." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 2 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "What is the behind insight in selecting a specific formula of $\\beta$ as $w_+^2-w_-^2$?\n\nDoes the theoretical analysis hold in the more general $n$-sample case, or what are the difficulties in this analysis?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The work seems to be solid.\n\nThe paper is polished very clearly.\n\nThe paper focuses on understanding the convergence of the GD algorithm, breaking through the limitations of traditional smoothness analysis (the step-size exceeds the threshold of $2/L$), and extending previous conclusions to more complex scenarios (from subquadratic growth objective functions to quadratic growth objective functions)." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper explores the Edge of Stability (EoS) phenomenon. They study GD and focus on a constant step-size exceeding the typical threshold of $2/L$, As the contributions, they observe that EoS occurs even when the loss $l$ is quadratic under proper conditions while the existing works require a subquadratic $l $. Due to the close relationship between the quadratic $l$ and the depth-2 diagonal linear network, the findings provide some explanations for the implicit bias of diagonal linear networks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Does the conclusion of the more general $n$-sample case in the paper apply to SGD? Due to the importance of stochastic optimization, I expect to see similar conclusions under stochastic optimization as well.\n\nI am sorry I am not very familiar with this topic. I will revise my score based on the comments of other more senior reviewers." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024criteria,\ntitle={Criteria and Bias of Parameterized Linear Regression under Edge of Stability Regime},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=0jJ94VVgzi},\nnote={under review}\n}" }, "abstract": { "value": "Classical optimization theory requires a small step-size for gradient-based methods to converge. Nevertheless, recent findings (Cohen et al., 2021) challenge the traditional idea by empirically demonstrating Gradient Descent (GD) converges even when the step-size $\\eta$ exceeds the threshold of $2/L$, where $L$ is the global smooth constant. This is usually known as the \\emph{Edge of Stability} (EoS) phenomenon. A widely held belief suggests that an objective function with subquadratic growth plays an important role in incurring EoS. In this paper, we provide a more comprehensive answer by considering the task of finding linear interpolator $\\beta \\in \\mathbb{R}^{d}$ for regression with loss function $l(\\cdot)$, where $\\beta$ admits parameterization as $\\beta = w^2_{+} - w^2_{-}$. Contrary to the previous work that suggests a subquadratic $l$ is necessary for EoS, our novel finding reveals that EoS occurs even when $l$ is quadratic under proper conditions. This argument is made rigorous by both empirical and theoretical evidence, demonstrating the GD trajectory converges to a linear interpolator in a non-asymptotic way. Moreover, the model under quadratic $l$, also known as a depth-$2$ \\emph{diagonal linear network}, remains largely unexplored under the EoS regime. Our analysis then sheds some new light on the implicit bias of diagonal linear networks when a larger step-size is employed, enriching the understanding of EoS on more practical models." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Edge of Stability", "gradient descent", "implicit bias" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/0cfb563bfb4a71e63523c2b1abec9c390548d930.pdf" }, "presentation": null, "primary_area": { "value": "optimization" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Criteria and Bias of Parameterized Linear Regression under Edge of Stability Regime" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "N/A" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The reviewer likes the integration of retrieval-based classification with CLIP-based scores to address the domain shift issues between masked images and natural images. It clearly improves the model's ability to recognize unseen classes without additional training.\n\n- The paper's approach to construct a feature database from widely available paired image-text data is interesting. This setup enables adaptability without requiring pixel-level annotations.\n\n- The paper is well-organized and well-written." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents an approach for open vocabulary panoptic segmentation by combining retrieval-based classification with standard image segmentation. In particular, the authors introduce a retrieval-augmented segmentation method that utilizes a database of paired image-text features. During inference to address the challenge of domain shift between masked and natural images, the model retrieves relevant features from this database using masked segment features from the input image as queries. This retrieval-based score is combined with scores from a vision-language model (CLIP) to enhance classification accuracy for unseen classes." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The reviewer feels that the retrieval-based classification relies heavily on the quality and diversity of the feature database constructed from paired image-text data. If the database lacks sufficient variety or coverage, the method may struggle to classify certain unseen classes accurately, particularly in real-world scenarios with a wide range of objects.\n\n- Further, the reviewer observed that the method uses Grounding DINO and SAM for generating masks in the training-free setup. However, SAM can produce suboptimal masks without human input which may degrade segmentation accuracy. This dependence on mask quality can limit the method’s effectiveness in fully automated settings.\n\n- The authors may want to include methods such as ODISE for a more comprehensive analysis." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. Could the authors elaborate on how their method generalizes to completely unseen classes that are not represented in the feature database?\n2. As the number of classes in the feature database grows, how does the retrieval process scale in terms of computational resources and accuracy?\n3. Are there any plans to compare the proposed method with other leading approaches in the field to contextualize the improvements?\n4. The paper mentions that the quality of mask proposal generation is crucial. Could the authors provide more details on how variations in mask quality affect the final segmentation results?\n5. Is there potential to integrate this method with other modalities, such as depth information or video data, to further improve performance?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The paper introduces a creative solution to the open vocabulary panoptic segmentation problem by combining retrieval-based classification with CLIP, which is an original approach not commonly seen in the literature.\n2. The paper is well-structured, with clear explanations of the methodology." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents a novel approach to address the challenge of segmenting arbitrary classes in images, a task known as open vocabulary panoptic segmentation. The authors propose a retrieval-augmented method that leverages a masked segment feature database constructed from image-text pairs. During inference, the system uses masked segment features from the input image to retrieve similar features and class labels from the database, combining these retrieval-based classification scores with CLIP-based scores to produce the final output. The method is evaluated on the ADE20k dataset and shows significant improvements over the baseline, particularly when fine-tuned on the COCO dataset, with absolute improvements of +4.5 PQ, +2.5 mAP, and +10.0 mIoU." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. While the paper demonstrates improvements over the baseline, it does not provide a direct comparison with other state-of-the-art methods in the field, which could provide additional context for the significance of the results.\n2. The discussion on how the proposed method generalizes to unseen classes could be expanded, as this is a critical aspect of open vocabulary segmentation.\n3. The paper could further discuss the limitations of the retrieval-augmented approach, especially regarding the reliance on the quality of the feature database and the potential scalability issues as the number of classes increases." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Please refer to the weaknesses. I think the current version is not ready for publication. More experiment results are expected." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "Applying Retrieval Augmentation to vision tasks is a promising direction. The proposed way of constructing a database is interesting." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper enhances open-vocabulary panoptic segmentation by leveraging retrieval augmentation to address the challenges of classifying unseen objects. The authors propose a framework that integrates masked segment features with a retrieval-based method to improve performance for unseen classes. The model builds a feature database using paired image-text data and retrieves similar features during inference to classify masked segments. These retrieval-based scores are combined with CLIP-based scores to enhance accuracy. When applied to FC-CLIP, the proposed method demonstrates improvements in unseen classes on the ADE20k dataset." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. While the method builds on FC-CLIP, the authors do not provide an introduction to FC-CLIP, which makes the paper hard to follow during reading.\n2. The feature database should be introduced prior to discussing the retrieval method to improve the flow and clarity of the paper.\n3. Since retrieval augmentation is intended to be a more general approach, the paper would benefit from presenting a more generalized framework to reflect its broader applicability.\n4. The method of constructing the feature database itself serves as a strong baseline. How does the performance of the proposed retrieval-augmentation approach compare to Grounding DINO?\n5. The paper lacks essential evaluations (the method is only evaluated on a single dataset with a single base model) and ablation studies." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "There is no ethics concerns." }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. What is the difference between IV Classification and OOV Classification cia CLIP in cross-dataset panoptic segmentation? What is the significance of this distinction? From Figure 1, it appears that the former only differs from the latter by including a linear projection.\n2. What is the fallback dataset, and how the author build it?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The paper explains related concepts clearly and details the methodology comprehensively, making the overall article easy to understand." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a retrieval-based method to enhance the performance of open vocabulary panoptic segmentation by constructing a feature database from paired image-text data. During inference, the model uses masked segment features from the input image to query the database for similar features and associated class labels, which are then combined with CLIP-based scores. This approach leads to improvements in Panoptic Quality in both training-free and cross-dataset settings." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The novelty of the paper is limited, primarily building upon the feature retrieval idea from Gui et al.[1]. Compared to Gui et al. [1]., the main modifications only include using a single CLIP backbone instead of two backbone. Please explain how these contributions can meet the strict requirements of top-level conferences.\n\n2. The authors use open vocabulary object detection combined with SAM to build the feature database, which limits the model's performance to the capabilities of the object detection component. Please explain how to handle classes that are not included in both the feature database and the fallback dataset during inference, or discuss the limitations of their approach for truly open-vocabulary scenarios.\n\n3. The definitions of IV Classification and OOV Classification are confusing. Why is it considered that the segment features and text embedding after linear projection in Figure 1 are equivalent to IV Classification? Please provide a more detailed explanation of the distinction between these two classifications and why the linear projection is significant for IV Classification.\n\n4. The experimental section lacks a critical component: comparisons with state-of-the-art methods, such as Gui et al. [1]., HIPIE [2], ODISE [3], OPSNet [4]. Please explain why these specific comparisons are not included and how your method compares theoretically to these state-of-the-art approaches.\n\n5. How does this method perform on open vocabulary semantic segmentation tasks, such as testing on ADE20K-847, ADE20K-150, Pascal Context-459.\n\n6. The paper claims to achieve performance improvement by utilizing a completely different dataset with only image level annotations. However, using the ADE20K training set to construct a feature database and evaluating it on the ADE20K validation set in the experiment lacks persuasiveness for open vocabulary. Please clarify how to ensure the open vocabulary nature when using the same dataset for both feature database construction and evaluation.\n\n7. There is irrelevant content in the lower-left corner of Figure 2. Please redraw the figure and ensure that the image is complete and free from irrelevant content\n\n\nreference:\n\n[1] Zhongrui Gui, Shuyang Sun, Runjia Li, Jianhao Yuan, Zhaochong An, Karsten Roth, Ameya Prabhu, and Philip Torr. knn-clip: Retrieval enables training-free segmentation on continually expanding large vocabularies, 2024. URL https://arxiv.org/abs/2404.09447.\n\n[2] Wang X, Li S, Kallidromitis K, et al. Hierarchical open-vocabulary universal image segmentation[J]. Advances in Neural Information Processing Systems, 2024, 36.\n\n[3] Xu J, Liu S, Vahdat A, et al. Open-vocabulary panoptic segmentation with text-to-image diffusion models[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023: 2955-2966.\n\n[4] Chen X, Li S, Lim S N, et al. Open-vocabulary panoptic segmentation with embedding modulation[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023: 1141-1150." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024open,\ntitle={Open Vocabulary Panoptic Segmentation With Retrieval Augmentation},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=0jUeqlQxMi},\nnote={under review}\n}" }, "abstract": { "value": "Given an input image and set of class names, panoptic segmentation aims to label each pixel in an image with class labels and instance labels. In comparison, Open Vocabulary Panoptic Segmentation aims to facilitate the segmentation of arbitrary classes according to user input. The challenge is that a panoptic segmentation system trained on a particular dataset typically does not generalize well to unseen classes beyond the training data. In this work, we propose a retrieval-augmented panoptic segmentation method that improves the performance of unseen classes. In particular, we construct a masked segment feature database using paired image-text data. At inference time, we use masked segment features from the input image as query keys to retrieve similar features and associated class labels from the database. Classification scores for the masked segment are assigned based on the similarity between query features and retrieved features. The retrieval-based classification scores are combined with CLIP-based scores to produce the final output. We incorporate our solution with a previous SOTA method (FC-CLIP). When trained on COCO, the proposed method demonstrates 30.9 PQ, 19.3 mAP, 44.0 mIoU on the ADE20k dataset, achieving +4.5 PQ, +2.5 mAP, +10.0 mIoU absolute improvement over the baseline." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Panoptic Segmentation", "Open Vocabulary", "Retrieval Augmentation" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/5778cf98bbda7a07ddd88560ada1fbb2c6994d19.pdf" }, "presentation": null, "primary_area": { "value": "applications to computer vision, audio, language, and other modalities" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Open Vocabulary Panoptic Segmentation With Retrieval Augmentation" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]