This repository contains the official implementation of Selective Entropy Regularization (SIREN), introduced in our paper: Rethinking Entropy Regularization in Large Reasoning Models.

SIREN addresses the issue of entropy collapse in Reinforcement Learning with Verifiable Reasoning (RLVR) when applying naive entropy regularization to large reasoning models. Built upon the veRL framework, our implementation introduces key modifications to entropy computation, aggregation, and the overall training objective.

We recommend creating a clean conda environment to avoid dependency conflicts.

conda create -n siren python=3.10
conda activate siren
pip install -r requirements.txt

# install verl
cd verl
pip install -e .

huggingface-cli download --repo-type dataset --resume-download Elliott/Openr1-Math-46k-8192 --local-dir data

We provide example scripts for both training and evaluation.

# training
bash exp_scripts/siren.sh

# evaluation
bash exp_scripts/eval.sh

• The training script (siren.sh) contains default hyperparameters and can be customized according to your experimental setup.

We thank the open-source communities behind the following projects for their valuable contributions:

• Frameworks: veRL, vLLM , Math-Verify
• Datasets: MATH, NuminaMath, OpenR1-Math-220k
• Backbones: Qwen2.5-Math, Llama-3.1

If you find our work useful in your research, please consider citing:

@misc{jiang2025rethinkingentropyregularizationlarge,
      title={Rethinking Entropy Regularization in Large Reasoning Models}, 
      author={Yuxian Jiang and Yafu Li and Guanxu Chen and Dongrui Liu and Yu Cheng and Jing Shao},
      year={2025},
      eprint={2509.25133},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2509.25133}, 
}