E-GRPO: High Entropy Steps Drive Effective Reinforcement Learning for Flow Models

This repository contains the weights for E-GRPO, as presented in the paper E-GRPO: High Entropy Steps Drive Effective Reinforcement Learning for Flow Models.

Introduction

E-GRPO (Entropy-Guided Group Relative Policy Optimization) is a reinforcement learning approach designed to enhance flow-matching models for human preference alignment. The key insight is that high-entropy denoising steps are more critical for policy optimization. The authors propose a merging-step strategy that focuses training on these important steps, leading to more efficient and effective exploration compared to standard SDE or ODE sampling methods.

Resources

• Paper: E-GRPO: High Entropy Steps Drive Effective Reinforcement Learning for Flow Models
• Code: GitHub - shengjun-zhang/VisualGRPO

Citation

If you find this work helpful for your research, please consider citing:

@article{zhang2025egrpo,
  title={E-GRPO: High Entropy Steps Drive Effective Reinforcement Learning for Flow Models},
  author={Zhang, Shengjun and Zhang, Zhang and Dai, Chensheng and Duan, Yueqi},
  journal={arXiv preprint arXiv:2601.00423},
  year={2025}
}