MedExAgent trains LLM agents to question patients, order exams, and diagnose in noisy clinical settings
Researchers formalize clinical diagnosis as a decision process with questioning, exam ordering, and diagnosis actions, then train an agent that balances accuracy against exam cost and patient discomfort.
In plain words
The researchers built a training environment where an AI agent must talk to a simulated patient, decide which medical tests to order, and then give a diagnosis, while dealing with incomplete or messy information. They trained their agent, MedExAgent, in two stages so it diagnoses accurately without ordering unnecessary or uncomfortable tests.
What was announced
The authors argue that existing benchmarks for medical large language models oversimplify diagnosis by reducing it to single-turn question answering, noise-free conversations, or sequential exam ordering, ignoring the interactive and uncertain character of real clinical work. Their environment casts diagnosis as a Partially Observable Markov Decision Process with three action types: questioning the patient, ordering medical exams as tool calls, and issuing a diagnosis. A noise model adds seven patient noise types and three exam noise types to simulate the incomplete, unreliable information doctors face.
MedExAgent is trained through a two-stage pipeline. First, supervised finetuning on synthetic conversations structured after the Calgary-Cambridge clinical interview model; second, DAPO reinforcement learning optimizes a composite reward covering diagnostic accuracy, tool call quality, and exam cost, where cost includes financial expense and patient discomfort. In experiments and ablation studies reported in the abstract, the authors state the agent achieves diagnostic performance comparable to larger models while keeping examination strategies cost-efficient. These results are self-reported and not independently verified.
What it does not show
- The evidence is the paper's abstract only; experimental details, baselines, and datasets are not visible, and the reported performance claims are the authors' own and not independently confirmed.
Sources
Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 20 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.