Research map / Reinforcement learning / RL from feedback
MedLoc-R1: Performance-Aware Curriculum Reward Scheduling for GRPO-Based Medical Visual Grounding
Guangjing Yang, Ziyuan Qin, Chaoran Zhang, Chenlin Du, Jinglin Wang, Wanran Sun, Zhenyu Zhang, Bing Ji, Qicheng Lao
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 |
|---|---|---|---|
| replace fixed IoU threshold 0.5 reward with performance-aware curriculum threshold | GRPO Qwen2.5-VL-3B · HEEL | A@0.8: 4.17 -> 47.35 | Improved |
| replace fixed IoU threshold 0.5 reward with performance-aware curriculum threshold | GRPO Qwen2.5-VL-3B · HAM10000 | A@0.5: 64.65 -> 94.46 | Improved |
| use performance-aware schedule instead of progress-only V-Triune schedule | GRPO Qwen2.5-VL-3B · TN3K | A@0.5: 43.50 -> 66.18 | Improved |
| use curriculum binary thresholds instead of continuous IoU reward | GRPO Qwen2.5-VL-3B · HEEL | A@0.8: 11.49 -> 47.35 | Improved |
| compare MedLoc-R1 to SFT | Qwen2.5-VL-3B · TN3K | A@0.8: 28.71 -> 29.60 | Improved |
| scale model from 3B to 7B/32B | MedLoc-R1 · TN3K | A@0.8: 29.60 -> 27.02 (7B), 27.52 (32B) | Mixed |
| scale model from 3B to 7B | MedLoc-R1 · HEEL | A@0.8: 47.35 -> 57.83 | Improved |
| add reward-check condition to threshold update criterion (ablation) | MedLoc-R1 · HAM10000 | A@0.5: 82.33 -> 94.96 | Improved |
| add IoU margin check to update criterion (ablation) | MedLoc-R1 · HAM10000 | A@0.5: 90.86 -> 94.96 | Improved |
| add stability (reward std) check to update criterion (ablation) | MedLoc-R1 · HAM10000 | A@0.5: 89.72 -> 94.96 | Improved |
| use dynamic step size/P/S instead of fixed aggressive setting (ablation) | MedLoc-R1 · HAM10000 | A@0.5: 71.54 -> 94.96 | Improved |
| use dynamic step size instead of identical constant step size (ablation) | MedLoc-R1 · HAM10000 | A@0.5: 79.64 -> 94.96 | Improved |
Closest papers by meaning
- MedReasoner: Reinforcement Learning Drives Reasoning Grounding from Clinical Thought to Pixel-Level Precision (AAAI 2026 · similarity 0.88)
- MedSG-Bench: A Benchmark for Medical Image Sequences Grounding (NeurIPS 2025 · similarity 0.86)
- MedEyes: Learning Dynamic Visual Focus for Medical Progressive Diagnosis (AAAI 2026 · similarity 0.86)
- QoQ-Med: Building Multimodal Clinical Foundation Models with Domain-Aware GRPO Training (NeurIPS 2025 · similarity 0.86)
- MedVR: Annotation-Free Medical Visual Reasoning via Agentic Reinforcement Learning (ICLR 2026 · similarity 0.86)
- Start Small, Think Big: Curriculum-based Relative Policy Optimization for Visual Grounding (AAAI 2026 · similarity 0.85)
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.
