MedAD-R1

Consistency-Reinforced Policy Optimization for Interpretable Medical Anomaly Detection

Access. Model access requests are reviewed while the paper is under review. We plan an open release after acceptance, subject to applicable permissions.

Model overview

MedAD-R1 is a compact multimodal model for interpretable medical anomaly detection. It answers medical-image questions while producing structured reasoning that is trained to remain grounded in the image and consistent with the final answer.

The model is trained on MedAD-38K, which contains 38,422 medical images and 130,146 VQA instances. Training combines supervised Cognitive Injection with Consistency Group Relative Policy Optimization (Con-GRPO).

Reported performance

Model / training	MedAD-38K accuracy	Reasoning-answer consistency	External accuracy
Qwen3.5-0.8B, zero-shot	76.43	92.12	68.12
Qwen3.5-0.8B, SFT only	90.89	94.35	88.36
MedAD-R1, SFT + Con-GRPO	95.12	97.26	90.09

See the paper for the full evaluation protocol and baseline comparison. This research checkpoint is intended for research use and has not been validated for clinical diagnosis or patient care.

Project resources

Resource	Links
Paper	MedAD-R1: Consistency-Reinforced Policy Optimization for Interpretable Medical Anomaly Detection
Code	GitHub
Dataset: MedAD-38K	Hugging Face · ModelScope
Model: MedAD-R1	Hugging Face · ModelScope

Citation

@article{zhang2026medadr1,
  title   = {MedAD-R1: Consistency-Reinforced Policy Optimization for Interpretable Medical Anomaly Detection},
  author  = {Zhang, Haitao and Wang, Yingying and Wang, Jiaxiang and Xu, Haote and Zhang, Hongyang and Chen, Yirong and Huang, Yue and Ding, Xinghao},
  journal = {arXiv preprint arXiv:2602.01081},
  year    = {2026}
}