This is the official implmentation of MedGround-R1, which incorporates GRPO and MedCLIP semantic reward. Without any cold start sft or cot annotation, MedGroud-R1 achieves sota on three public medical grounding benchmarks.

• upload checkpoints and logs.
• 09/24 upload dataset and log for MS-CXR
• 09/01 miccai'25 best paper shortlist 😊
• 08/10 miccai'25 splotlight
• 07/01 camera ready paper upload to arxiv
• 05/10 miccai'25 early accept

The author is committed to fully open-access and transparency, all checkpoints will be uploaded before 10/08 holiday.

1. Installation for VLM-R1: Our repo is based on VLM-R1, please refer to vlm-r1 for installation guidance.

2. Installation for MedCLIP: Please refer to medclip for installation guidance.

3. Training: Using 8 H100 80G:

cd open-r1-multimodal/src/open-r1
bash run_grpo_rec.sh

4. Evaluate:

python eval/test_rec_r1_med_qwen25.py

5. logs, checkpoints, predictions

@article{xu2025medground,
  title={Medground-r1: Advancing medical image grounding via spatial-semantic rewarded group relative policy optimization},
  author={Xu, Huihui and Nie, Yuanpeng and Wang, Hualiang and Chen, Ying and Li, Wei and Ning, Junzhi and Liu, Lihao and Wang, Hongqiu and Zhu, Lei and Liu, Jiyao and others},
  journal={arXiv preprint arXiv:2507.02994},
  year={2025}
}

• Thanks VLM-R1 for the grpo implementation.
• Thanks MedCLIP for the medclip implementation.