Analyzing and Improving Fine-grained Preference Optimization in Medical LVLMs
FIRE-MPO is a new fine-grained alignment framework for medical Vision-Language Models that solves issues with existing methods like DPO. It uses a bidirectional token-wise KL regularizer and a visual grounding objective to ensure the model focuses on clinically accurate information and links its answers directly to specific visual evidence, leading to significant performance gains in medical tasks.