Applying Reinforcement Learning to Vision-Language Models for Mathematical Reasoning

A minimal implementation of PPO-based Reinforcement Learning for vision-language models, focusing on mathematical reasoning tasks.

🚧 Work in progress...

🙏 Acknowledgements

This implementation is inspired by the MAYE framework described in "Rethinking RL Scaling for Vision Language Models: A Transparent, From-Scratch Framework and Comprehensive Evaluation Scheme".

@article{ma2025maye,
  title={Rethinking RL Scaling for Vision Language Models: A Transparent, From-Scratch Framework and Comprehensive Evaluation Scheme},
  author={Ma, Yan and Chern, Steffi and Shen, Xuyang and Zhong, Yiran and Liu, Pengfei},
  journal={arXiv preprint arXiv:2504.02587},
  year={2025},
}