GMAI-VL-R1

🔬 Official codebase for the paper:
GMAI-VL-R1: Harnessing Reinforcement Learning for Multimodal Medical Reasoning
arXiv: 2504.01886 · Yanzhou Su, Tianbin Li, Jiyao Liu, Chenglong Ma, Junzhi Ning, Cheng Tang, Sibo Ju, Jin Ye, Pengcheng Chen, Ming Hu, et al.

---

📘 Introduction

This repository contains the implementation of GMAI-VL-R1, a framework for multimodal medical reasoning powered by reinforcement learning.

We aim to explore how large vision-language models (VLMs) can improve medical visual understanding and reasoning with rule-based reward design and high-quality medical datasets.

Our method incorporates recent advancements in reinforcement learning for vision-language models and adapts them to the challenging domain of medical imaging.

---

🧠 Built Upon

Our work builds upon and integrates ideas from several open-source projects:

• MM-Eureka-V0 (FanqingM/MM-Eureka-V0)
• open-r1-multimodal (EvolvingLMMs-Lab/open-r1-multimodal)
• MM-EUREKA (ModalMinds/MM-EUREKA)

We thank the authors of these projects for their valuable contributions to the community.

---

📊 Dataset

We use a high-quality medical reasoning dataset:

📁 Dataset Name: GMAI-Reasoning10K
📄 Format: JSONL
📷 Modalities: X-ray, CT, MRI, OCT, Ultrasound, etc.
🔢 Samples: 10,000 high-quality multiple-choice questions designed for reinforcement learning training
📚 Source: Aggregated from 95 public medical datasets (e.g., Kaggle, GrandChallenge)

---

📜 Citation

If you find our work helpful, please consider citing our paper:

@article{su2025gmai,
  title={Gmai-vl-r1: Harnessing reinforcement learning for multimodal medical reasoning},
  author={Su, Yanzhou and Li, Tianbin and Liu, Jiyao and Ma, Chenglong and Ning, Junzhi and Tang, Cheng and Ju, Sibo and Ye, Jin and Chen, Pengcheng and Hu, Ming and others},
  journal={arXiv preprint arXiv:2504.01886},
  year={2025}
}