Reproduce “Improving Medical Visual Reinforcement Fine-Tuning via Perception and Reasoning Augmentation” on OpenResearch

About this paper

Guangjing Yang, ZhangYuan Yu, Ziyuan Qin, Xinyuan Song, Huahui Yi, Qingbo Kang, Jun Gao, Yiyue Li and 2 more

This research presents VRFT-Aug, a visual reinforcement fine-tuning framework that enhances Large Vision-Language Models' (LVLMs) perception and reasoning capabilities for medical imaging tasks. It achieves improved accuracy and training stability by employing knowledge-augmented prompts, cross-task training for localization, and medical-specific reward schemes on various medical datasets.

Key insights

• Explicitly embedding medical domain knowledge into prompts and implicitly training for localization significantly enhances LVLM perception and overall task performance in medical visual recognition.
• Reward shaping is critical for effective medical V-RFT; discouraging excessive 'recitation' of prior knowledge improves model flexibility and generalization, while fuzzy rewards are vital for stable learning in fine-grained ordinal classification.
• The Multi-Grade Fuzzy Reward Scheme successfully mitigates the sparse reward problem in medical grading tasks, allowing the model to learn effectively from 'near-correct' predictions and develop more robust reasoning patterns.

Summary by alphaXiv.