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

2026-07-03

AI Summary

Current large language models in medical diagnosis predominantly rely on passive reasoning, failing to emulate the active, iterative evidence-gathering process characteristic of clinical practice. This work proposes the RLVR training framework coupled with RAGES, a high-fidelity clinical simulator, to formulate diagnosis as an iterative evidence-seeking task within a closed-loop system. By integrating reinforcement learning, a verifiable reward mechanism, and retrieval-augmented generation (RAG), the approach enables models to autonomously plan diagnostic tests and dynamically acquire relevant evidence. Experimental results demonstrate that the method achieves performance on par with substantially larger or specialized reasoning models across multiple datasets. Moreover, clinical feedback generated by RAGES exhibits significantly higher biological plausibility compared to baseline large language models.

Abstract

Recent reasoning-centric Large Language Models (LLMs) have made significant strides, yet they predominantly operate on a passive-inference pattern that assumes complete information. In contrast, real-world clinical intelligence is inherently an iterative investigative process requiring strategic evidence acquisition. To bridge this gap, we formalize medical diagnosis as an Iterative Evidence-Seeking Task. We leverage Reinforcement Learning with Verifiable Rewards (RLVR) to elicit intrinsic reasoning within a closed-loop environment, guided by a novel suite of rewards that enforce diagnostic precision and examination consistency. To facilitate this, we introduce the Retrieval-Augmented Generation-based Examination Simulator (RAGES), a high-fidelity clinical oracle that provides realistic, knowledge-grounded follow-up evidence. Empirical results across diverse datasets demonstrate that our framework enables LLMs to transition from passive responders to autonomous assistants. Notably, our model demonstrates comparable performance to larger and reasoning-enhanced baselines, while RAGES proves superior to vanilla LLMs in generating biologically plausible clinical feedback.

Problem

• Evidence-Seeking
• Diagnostic Reasoning
• Large Language Models
• Reinforcement Learning
• Clinical Intelligence

Innovation

• Reinforcement Learning with Verifiable Rewards
• Iterative Evidence-Seeking
• Retrieval-Augmented Generation
• Clinical Reasoning
• Closed-loop Diagnostic Framework