An AI research team has trained language models to behave more like clinicians — asking follow-up questions before reaching a diagnosis, rather than pattern-matching on whatever information arrives first.
The paper, posted to arXiv, reframes medical diagnosis as an "Iterative Evidence-Seeking Task." The researchers applied Reinforcement Learning with Verifiable Rewards (RLVR) — a training method that rewards models for reaching correct, internally consistent conclusions — inside a closed feedback loop. To give the model something realistic to query, they built a retrieval-augmented simulator called RAGES that generates clinically plausible follow-up findings, standing in for a real patient chart or lab system. Results across multiple datasets showed their smaller model matching or approaching the performance of larger, reasoning-tuned baselines.
Most large language models are built around a passive assumption: all the relevant context arrives in the prompt, and the model's job is to process it. Clinical medicine doesn't work that way — a doctor orders tests, waits for results, and revises. A model trained to seek evidence rather than infer from silence is structurally better suited to settings where incomplete information is the default, not the exception. That gap has been a persistent criticism of AI diagnostic tools.
The catch is that RAGES is still a simulator, not a real EHR or lab system — and the distance between a "biologically plausible" synthetic response and a real patient's messy chart is exactly where these systems tend to break in deployment.