CARing Framework Improves Medical Diagnosis Prediction via Coverage-Aware Reasoning

CARing addresses LLM limitations in next-visit diagnosis prediction by utilizing specialized medical tokens and a coverage-aware reward mechanism. This prevents the model from over-optimizing for a few high-probability diagnoses while neglecting the full set of valid clinical possibilities.