Reinforcement Learning for Evidence-Seeking Diagnostic Reasoning with Large Language Models

This paper formulates medical diagnosis as an Iterative Evidence-Seeking Task rather than a one-shot inference problem. It applies Reinforcement Learning with Verifiable Rewards (RLVR) to train language models to acquire follow-up examinations, using rewards for diagnostic precision and examination consistency. The authors introduce RAGES, a Retrieval-Augmented Generation-based Examination Simulator intended to provide realistic, knowledge-grounded clinical feedback. According to the abstract, experiments across diverse datasets show performance comparable to larger and reasoning-enhanced baselines, while RAGES produces more biologically plausible feedback than vanilla LLMs. The abstract does not provide dataset names or quantitative scores.