Reinforcement Learning for Evidence-Seeking Diagnostic Reasoning with Large Language Models
06:00 · July 7, 2026 · arXiv cs.AI RSS
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.
Summary
Recent reasoning-centric large language models have advanced chain-of-thought capabilities, yet most still rely on a passive-inference pattern that assumes all necessary information is already present. In clinical practice, diagnosis is an active, iterative process: clinicians begin with incomplete observations and must strategically request further examinations to refine hypotheses under uncertainty. This paper addresses that gap by formalizing pathological diagnosis as an Iterative Evidence-Seeking Task, in which an LLM must generate differential hypotheses and propose auxiliary tests within a closed-loop environment.
The training approach centers on Reinforcement Learning with Verifiable Rewards (RLVR) implemented through the Group Relative Policy Optimization framework. A tri-factor reward structure guides the model: a format reward maintains structural coherence, a rank-sensitive diagnostic reward encourages precise yet comprehensive differential lists, and an examination consistency reward aligns proposed tests with biologically plausible requests. A higher-capacity LLM serves as a reasoning verifier to evaluate logical consistency without relying on subjective human judgment.
To supply realistic feedback, the authors introduce the Retrieval-Augmented Generation-based Examination Simulator (RAGES). This component functions as a knowledge-grounded clinical oracle that draws on a curated pathological corpus to deliver deterministic, biologically plausible laboratory or morphological results in response to model queries. The resulting loop enables the LLM to acquire information incrementally rather than issuing a single-turn prediction.
Empirical evaluation across multiple datasets shows that the resulting 7B-parameter model achieves diagnostic accuracy comparable to larger reasoning-enhanced baselines while improving differential accuracy. RAGES itself outperforms vanilla LLMs in generating clinically coherent follow-up evidence. The work thereby demonstrates a scalable route from static, fully observed benchmarks toward autonomous, evidence-seeking diagnostic assistants.
Why it matters
This research is highly relevant for Dutch AI researchers and health-tech enterprises developing autonomous clinical assistants. The use of RLVR and RAGES provides a novel, actionable methodology for creating more accurate, iterative, and verifiable medical AI systems, aligning with the EU's focus on robust healthcare AI.
