Research map / Video / Video-language understanding
MedGRPO: Multi-Task Reinforcement Learning for Heterogeneous Medical Video Understanding
Yuhao Su, Anwesa Choudhuri, Zhongpai Gao, Benjamin Planche, Van Nguyen Nguyen, Meng Zheng, Yuhan Shen, Arun Innanje, Terrence Chen, Ehsan Elhamifar, Ziyan Wu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add cross-dataset reward normalization to GRPO (ablation) | MedGRPO · MedVidBench | CVS accuracy: 0.020 -> 0.896 | Improved |
| add cross-dataset reward normalization to GRPO (ablation) | MedGRPO · MedVidBench | TAG mIoU@0.3: 0.004 -> 0.216 | Improved |
| add cross-dataset reward normalization to GRPO (ablation) | MedGRPO · MedVidBench | VS LLM judge: 3.805 -> 4.184 | No clear effect |
| add caption rewards (VS+RC) alongside grounding rewards (ablation) | MedGRPO · MedVidBench | TAG mIoU@0.3: 0.202 -> 0.216 | Improved |
| add caption rewards alongside grounding rewards (ablation) | MedGRPO · MedVidBench | STG mIoU: 0.193 -> 0.202 | Improved |
| add caption rewards alongside grounding rewards (ablation) | MedGRPO · MedVidBench | DVC F1: 0.225 -> 0.214 | Hurt |
| add medical LLM judge to caption reward (ablation) | MedGRPO (VS+RC) · MedVidBench | RC LLM judge: 2.984 -> 3.235 | Improved |
| add medical LLM judge to caption reward (ablation) | MedGRPO (VS+RC) · MedVidBench | VS LLM judge: 3.733 -> 3.824 | Improved |
| apply MedGRPO RL on top of SFT | Qwen2.5VL-7B SFT · MedVidBench | VS LLM judge: 3.596 -> 4.184 | Improved |
| apply MedGRPO RL on top of SFT | Qwen2.5VL-7B SFT · MedVidBench | NAP accuracy: 0.442 -> 0.405 | Hurt |
| apply MedGRPO RL on top of SFT | Qwen2.5VL-7B SFT · MedVidBench | CVS accuracy: 0.894 -> 0.896 | No clear effect |
| SFT on MedVidBench vs off-the-shelf Qwen2.5VL-7B | Qwen2.5-VL-7B · MedVidBench | CVS accuracy: 0.105 -> 0.894 | Improved |
Closest papers by meaning
- MedReasoner: Reinforcement Learning Drives Reasoning Grounding from Clinical Thought to Pixel-Level Precision (AAAI 2026 · similarity 0.89)
- QoQ-Med: Building Multimodal Clinical Foundation Models with Domain-Aware GRPO Training (NeurIPS 2025 · similarity 0.89)
- Token-Sparse Medical Multimodal Reasoning via Dual-Stream Reinforcement Learning (ICML 2026 · similarity 0.88)
- MedUnifier: Unifying Vision-and-Language Pre-training on Medical Data with Vision Generation Task using Discrete Visual Representations (CVPR 2025 · similarity 0.88)
- MedMO: Grounding and Understanding Multimodal Large Language Model for Medical Images (CVPR 2026 · similarity 0.88)
- MedVR: Annotation-Free Medical Visual Reasoning via Agentic Reinforcement Learning (ICLR 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
