Multimodal Knowledge Transfer to LLMs via Preference Optimization for Biomedical Applications

Summary

Wang et al. introduce MINT, a framework that transfers knowledge from multimodal biomedical models to unimodal large decoder models through preference optimization. Applied to rare-disease prediction and tissue-type classification, MINT improves task performance while preserving general reasoning ability and offers a practical strategy for extending scarce multimodal biomedical knowledge to more flexible foundation models.