Papers
arxiv:2510.14545

Agentic Entropy-Balanced Policy Optimization

Published on Oct 16
· Submitted by KABI on Oct 17
#2 Paper of the day
Authors:
,
,
,
,
,
,
,
,
,
,
,

Abstract

AEPO, an agentic RL algorithm, addresses entropy-related challenges in web agent training, enhancing performance and stability across various datasets.

AI-generated summary

Recently, Agentic Reinforcement Learning (Agentic RL) has made significant progress in incentivizing the multi-turn, long-horizon tool-use capabilities of web agents. While mainstream agentic RL algorithms autonomously explore high-uncertainty tool-call steps under the guidance of entropy, excessive reliance on entropy signals can impose further constraints, leading to the training collapse. In this paper, we delve into the challenges caused by entropy and propose the Agentic Entropy-Balanced Policy Optimization (AEPO), an agentic RL algorithm designed to balance entropy in both the rollout and policy update phases. AEPO comprises two core components: (1) a dynamic entropy-balanced rollout mechanism that adaptively allocate global and branch sampling budget through entropy pre-monitoring, while imposing a branch penalty on consecutive high-entropy tool-call steps to prevent over-branching issues; and (2) Entropy-Balanced Policy Optimization that inserts a stop-gradient operation into the high-entropy clipping term to preserve and properly rescale gradients on high-entropy tokens, while incorporating entropy-aware advantage estimation to prioritize learning on high-uncertainty tokens. Results across 14 challenging datasets show that AEPO consistently outperforms 7 mainstream RL algorithms. With just 1K RL samples, Qwen3-14B with AEPO achieves impressive results: 47.6% on GAIA, 11.2% on Humanity's Last Exam, and 43.0% on WebWalker for Pass@1; 65.0% on GAIA, 26.0% on Humanity's Last Exam, and 70.0% on WebWalker for Pass@5. Further analysis reveals that AEPO improves rollout sampling diversity while maintaining stable policy entropy, facilitating scalable web agent training.

Community

Paper author Paper submitter
edited 6 days ago

We propose Agentic Entropy-Balanced Policy Optimization (AEPO), an entropy-balanced agentic RL algorithm designed for training multi-turn web agents. AEPO focuses on balancing and rationalizing rollout branching and policy updates under the guidance of high-entropy tool calls, thereby achieving more stable RL training.

image

With just 1𝐾 RL samples, Qwen3-14B with AEPO achieves impressive results: 47.6% on GAIA, 11.2% on Humanity’s Last Exam, and 43.0% on WebWalkerQA for Pass@1; 65.0% on GAIA, 26.0% on Humanity’s Last Exam, and 70.0% on
WebWalkerQA for Pass@5.

Paper author Paper submitter
edited 6 days ago

🔧 All the code, datasets and model checkpoints of AEPO are fully open-sourced:

Github: https://github.com/dongguanting/ARPO

Models: https://huggingface.co/collections/dongguanting/aepo-68ef6832c99697ee03d5e1c7

🔥 Key Insights:

  1. We systematically reveal two entropy-driven issues inherent to agentic RL: "High-Entropy Rollout Collapse" and "High-Entropy Token Gradient Clipping" (as shown in the above figure). Through preliminary experiments, we quantify their impact on multi-turn web-agent training, offering empirical evidence for further research into entropy balancing.

  2. We propose a Dynamic Entropy-Balanced Rollout mechanism , which adaptively allocates rollout sampling budgets via entropy pre-monitoring, while imposing a branch penalty on consecutive high-entropy steps to prevent over-branching issues.

  3. We introduce Entropy-Balanced Policy Optimization , which intuitively integrates a stop-gradient operation into the high-entropy clipping term to preserve and rescale gradients on high-entropy tokens, while incorporating entropy-aware advantage estimation to prioritize learning on high-uncertainty tokens.

  4. Experiments on 14 challenging benchmarks demonstrate that AEPO consistently outperforms 7 mainstream RL algorithms in web agent training. With just 1𝐾 RL samples, Qwen3-14B with AEPO achieves impressive results: 47.6% on GAIA, 11.2% on Humanity’s Last Exam, and 43.0% on WebWalkerQA for Pass@1; 65.0% on GAIA, 26.0% on Humanity’s Last Exam, and 70.0% on WebWalkerQA for Pass@5.

✨ Two entropy-driven challenges:

image

🔥 Overview of AEPO:

image

This is an automated message from the Librarian Bot. I found the following papers similar to this paper.

The following papers were recommended by the Semantic Scholar API

Please give a thumbs up to this comment if you found it helpful!

If you want recommendations for any Paper on Hugging Face checkout this Space

You can directly ask Librarian Bot for paper recommendations by tagging it in a comment: @librarian-bot recommend

Sign up or log in to comment

Models citing this paper 3

Datasets citing this paper 3

Spaces citing this paper 0

No Space linking this paper

Cite arxiv.org/abs/2510.14545 in a Space README.md to link it from this page.

Collections including this paper 13