Agentic-R1: Distilled Dual-Strategy Reasoning

This repository hosts the Agentic-R1 model, an implementation of the paper Agentic-R1: Distilled Dual-Strategy Reasoning.

Code: https://github.com/StigLidu/DualDistill

Abstract

Current long chain-of-thought (long-CoT) models excel at mathematical reasoning but rely on slow and error-prone natural language traces. Tool-augmented agents address arithmetic via code execution, but often falter on complex logical tasks. We introduce a fine-tuning framework, DualDistill, that distills complementary reasoning strategies from multiple teachers into a unified student model. Using this approach, we train Agentic-R1, which dynamically selects the optimal strategy for each query, invoking tools for arithmetic and algorithmic problems, and using text-based reasoning for abstract ones. Our method improves accuracy across a range of tasks, including both computation-intensive and standard benchmarks, demonstrating the effectiveness of multi-strategy distillation in achieving robust and efficient reasoning.

Key Features

• Efficient Training: Integrates tool use into long-chain-of-thought (CoT) reasoning using only 4 × A6000 GPUs
• Unified Reasoning: Fuses heterogeneous reasoning traces from multiple teacher models into a single student model

Overview of DualDistill methodology

Datasets

Dataset	Description	Link
Training Set	Complete training dataset with teacher trajectories	🤗 HuggingFace
Test Set	Evaluation benchmarks	dataset/test/

Results

• Agentic-R1 demonstrates significant performance gains on DeepMath-L and Combinatorics300, where both complex reasoning and tool use are crucial for success.
• Agentic-R1-SD (Self-Distilled) further enhances performance through our self-distillation approach, consistently outperforming baseline models across nearly all evaluation tasks.

Quick Start

Installation

1. Clone the repository:

   git clone https://github.com/StigLidu/DualDistill.git
   cd DualDistill
   

2. Create environment (optional but recommended):

   conda create -n dualdistill python=3.11
   conda activate dualdistill
   

3. Install dependencies:

   pip install -r requirements.txt
   pip install flash-attn --no-build-isolation
   

Sample Usage

Here's how to perform inference with the Agentic-R1 model using the Hugging Face transformers library:

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "VanishD/Agentic-R1" # Or "VanishD/Agentic-R1-SD" for the self-distilled version

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16, # Use bfloat16 for better performance and memory if supported
    device_map="auto",
    trust_remote_code=True
).eval() # Set model to evaluation mode

# Prepare a simple user message
messages = [{"role": "user", "content": "What is 123 + 456?"}]

# Apply the chat template to format the prompt correctly for the model
# The `add_generation_prompt=True` adds the Assistant token to prompt the model for its response.
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

# Encode the prompt
input_ids = tokenizer(prompt, return_tensors="pt").input_ids.to(model.device)

# Generate response
output_ids = model.generate(
    input_ids,
    max_new_tokens=256,
    do_sample=True,
    temperature=0.7,
    top_p=0.95,
    eos_token_id=tokenizer.eos_token_id,
    pad_token_id=tokenizer.pad_token_id, # Often EOS token is used as PAD token for LLMs
)

# Decode and print the generated text, excluding the input prompt
generated_text = tokenizer.decode(output_ids[0][input_ids.shape[1]:], skip_special_tokens=True).strip()
print(f"Generated Text:
{generated_text}")

⚠️ Important Notes

• Code Execution Safety: The evaluation scripts execute model-generated code locally. Only use trusted models before execution.
• Inference Config: If you are using vLLM (a recent version) and encounter an error regarding the maximum context length. You may need to modify the model_max_length in tokenizer_config.json.
• Self-Distillation Warning: The self-distillation step requires sampling many trajectories and can be time-consuming.

License

This project is licensed under the MIT License - see the LICENSE file for details.

Acknowledgments

We thank the following open-source projects for their foundational contributions:

• OpenHands - Agent framework
• DeepMath-103K - Mathematical reasoning dataset
• vLLM - High-performance inference engine

Contact

For questions or support, please contact:

• Weihua Du: weihuad@cs.cmu.edu

Citation

If you find our work useful, please consider citing:

@article{du2025agentic,
  title={Agentic-R1: Distilled Dual-Strategy Reasoning},
  author={Du, Weihua and Aggarwal, Pranjal and Welleck, Sean and Yang, Yiming},
  journal={arXiv preprint arXiv:2507.05707},
  year={2025}
}