C-MIG: Multi-view Information Gain-based Retrieval-Augmented Generation for Clinical Diagnosis Reasoning

Researchers introduce C-MIG, a retrieval-augmented generation framework that improves clinical diagnosis reasoning by using multi-view information gain instead of binary reward signals. The method outperforms existing RAG-RL approaches on medical benchmarks by better capturing semantically relevant information and addressing credit assignment challenges in healthcare AI systems.

C-MIG represents a meaningful advancement in medical AI by solving a fundamental limitation of current retrieval-augmented generation systems combined with reinforcement learning. Traditional approaches rely on exact-match binary rewards, creating a brittle evaluation mechanism that discards valuable learning signals when clinical reasoning steps are semantically correct but linguistically different from reference answers. This architectural flaw particularly undermines performance in medical domains where clinical knowledge can be expressed multiple valid ways.

The framework addresses this through dual-perspective information gain estimation—examining both retrieved documents and document refinements—to simultaneously optimize what information the model retrieves and how it processes that information. This multi-dimensional reward structure enables more nuanced credit assignment across heterogeneous reasoning capabilities required for diagnosis.

The clinical AI market increasingly demands systems grounded in trustworthy medical evidence, as regulations tighten around healthcare AI accountability. C-MIG's demonstrated performance gains across both in-domain and out-of-domain medical benchmarks suggests practical improvements in generalization, critical for deployment across diverse clinical settings and patient populations. The multi-subquery retrieval strategy specifically designed for diagnostic scenarios indicates domain-aware engineering beyond generic language model optimization.

For healthcare technology developers and clinical AI researchers, C-MIG offers a methodological template for improving reward design in specialized domains. The framework's superior performance compared to general-purpose LLMs demonstrates that domain-specific architectural choices outperform scaling alone in medical applications, influencing investment priorities in clinical AI development.

• →C-MIG replaces binary rewards with multi-view information gain metrics, capturing semantically valid clinical reasoning that traditional exact-match systems discard.
• →The framework jointly optimizes retrieval and document refinement stages through complementary reward signals for better credit assignment.
• →Multi-subquery retrieval strategy specifically designed for clinical diagnosis improves knowledge recall coverage beyond generic RAG approaches.
• →Comprehensive testing shows C-MIG outperforms state-of-the-art general LLMs and competing RAG-RL methods on four medical benchmarks.
• →Architecture demonstrates domain-specific optimization in clinical AI yields better results than scaling general-purpose language models.