Research map / Reinforcement learning / RL from feedback
Regulating Anatomy-Aware Rewards via Trajectory-Integral Feedback for Volumetric Computed Tomography Analysis
Tianwei Lin, Zhongwei Qiu, Jie Cao, Jiang Liu, Wenjie Yan, Zhang Bo, Yu Zhong, Wenqiao Zhang, Yingda Xia, Ling Zhang
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 TIF-GRPO RL stage after SFT | Qwen3-VL-4B · CT-RATE-Report | entity F1: 18.73 -> 39.11 | Improved |
| add TIF-GRPO RL stage after SFT | Qwen3-VL-4B · AMOS-MM-Report | entity F1: 17.65 -> 30.47 | Improved |
| scale TIF-GRPO model from 4B to 8B | TIF-GRPO · CT-RATE-Report | entity F1: 39.11 -> 39.67 | No clear effect |
| replace ROUGE reward with trajectory-integral CABS reward (ablation) | GRPO · CT-RATE-Report | entity F1: 29.74 -> 39.11 | Improved |
| replace RadGraph reward with TIF reward (ablation) | GRPO · CT-RATE-Report | entity F1: 25.32 -> 39.11 | Improved |
| replace LLM rubric reward with TIF reward (ablation) | GRPO · CT-RATE-Report | entity F1: 32.73 -> 39.11 | Improved |
| use unbalanced running cost vs control effort weights (0.4/1.0) (ablation) | TIF-GRPO · CT-RATE-Report | entity F1: 39.11 -> 36.96 | Hurt |
| use unbalanced running cost vs control effort weights (1.0/0.4) (ablation) | TIF-GRPO · CT-RATE-Report | entity F1: 39.11 -> 37.32 | Hurt |
| swap GRPO for RLOO, ReMAX or Reinforce++ with surface rewards (ablation) | TIF-GRPO · CT-RATE-Report | entity F1: TIF-GRPO 39.11 vs best alternative 35.20 | Hurt |
| add TIF control on top of accuracy-reward GRPO (ablation) | GRPO w/ Acc Reward (AMOS-MM) · CT-RATE-MCQ + AMOS-MM-MCQ | total average accuracy: 63.91 -> 69.67 | Improved |
| add TIF control on top of accuracy-reward GRPO (ablation) | GRPO w/ Acc Reward (CT-RATE) · CT-RATE-MCQ + AMOS-MM-MCQ | total average accuracy: 63.59 -> 66.65 | Improved |
| apply TIF-GRPO to 2D chest X-ray report generation | Qwen3-VL · MIMIC-CXR-Report | F1: 25.3 -> 70.8 | Improved |
Closest papers by meaning
- 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)
- MedReasoner: Reinforcement Learning Drives Reasoning Grounding from Clinical Thought to Pixel-Level Precision (AAAI 2026 · similarity 0.86)
- DDx-TRACE: A Benchmark for Medical Diagnostic Trajectories in VLMs (NeurIPS 2026 · similarity 0.86)
- 3DMedAgent: Unified Perception-to-Understanding for 3D Medical Analysis (ICML 2026 · similarity 0.86)
- OralGPT-Plus: Learning to Use Visual Tools via Reinforcement Learning for Panoramic X-ray Analysis (CVPR 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.
