C-MIG: Multi-View Information Gain for Clinical Diagnosis RAG

A team of researchers has introduced C-MIG, a framework for retrieval-augmented generation based on Multi-view Information Gain aimed at enhancing clinical diagnosis reasoning. Current RAG approaches utilizing reinforcement learning rely on exact-match binary rewards, leading to two main problems: semantically relevant but non-verbatim steps do not earn rewards, and one-dimensional rewards fail to manage diverse reasoning. C-MIG evaluates information gain through two complementary perspectives—retrieved documents and document refinement—using a fixed reference model to direct both retrieval and refinement processes. Additionally, it features a strategy for multi-subquery retrieval augmentation, effectively tackling issues related to reward signal loss and credit assignment in clinical reasoning tasks.