LLMs Learn Evidence-Seeking for Medical Diagnosis via Reinforcement Learning.

Current Large Language Models (LLMs) often struggle with medical diagnosis because they typically operate on a passive-inference model, assuming all necessary information is already present. Real-world clinical diagnosis, however, is an iterative process requiring active evidence acquisition. This paper addresses this by formalizing medical diagnosis as an "Iterative Evidence-Seeking Task." The researchers propose a framework utilizing Reinforcement Learning with Verifiable Rewards (RLVR) to train LLMs for this task. This approach encourages intrinsic reasoning within a closed-loop environment, guided by new reward structures that emphasize diagnostic precision and consistent examination. To support this, they developed RAGES (Retrieval-Augmented Generation-based Examination Simulator), a high-fidelity clinical oracle that provides realistic, knowledge-grounded follow-up evidence. Empirical results across various datasets show that this framework transforms LLMs into autonomous assistants capable of actively seeking evidence. The model achieved performance comparable to much larger and more reasoning-enhanced baselines, with RAGES proving superior to vanilla LLMs in generating biologically plausible clinical feedback.