MedCEG: Reinforcing Verifiable Medical Reasoning with Critical Evidence Graph

The core application domain is medical artificial intelligence, specifically improving the reliability and trustworthiness of Large Language Models (LLMs) for clinical reasoning. MedCEG addresses the critical need for transparent and clinically valid reasoning in medical contexts, where illogical or invalid reasoning can have severe consequences. By explicitly supervising the reasoning process with Critical Evidence Graphs, the framework aims to prevent 'shortcuts' in reasoning that might lead to correct answers for the wrong reasons, thereby enhancing human-AI collaboration and educational utility in clinical settings.

A central innovation of MedCEG is the algorithmic construction and utilization of Critical Evidence Graphs (CEGs). These graphs serve as structured, explicit representations of high-quality verifiable reasoning pathways, externalizing the implicit logic of clinical narratives. The CEGs capture essential clinical entities, their relationships, and causal pathways, acting as a direct, non-learned reward signal in the reinforcement learning phase. This graph-based approach ensures that models learn to prioritize critical inferential steps and generate explanations that follow clinically valid trajectories.

The paper places significant emphasis on robust evaluation, not just of final answer accuracy, but also of the intrinsic quality and trustworthiness of the reasoning process itself. It introduces a composite Clinical Reasoning Procedure (CRP) reward, which integrates metrics for Node Coverage, Structural Correctness, and Chain Completeness. Beyond quantitative performance, the framework also addresses ethical considerations, positioning MedCEG as a clinical decision support tool that requires human verification, emphasizing transparency and avoiding hallucinations for patient safety.