Research map / Multi-agent systems and games / LLM agents

MedExAgent: Training LLM Agents to Ask, Examine, and Diagnose in Noisy Clinical Environments

NeurIPS 2026 · Multi-agent systems and games › LLM agents · approach: Agent data synthesis and behavior studies · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Formalizes noisy, cost-aware interactive clinical diagnosis as a POMDP and trains an 8B Meditron3-based agent via Calgary-Cambridge-structured SFT followed by DAPO with a composite diagnosis/tool-match/exam-cost reward.

Problem

Existing medical LLM benchmarks and diagnosis agents reduce clinical diagnosis to single-turn QA, noise-free conversation, or sequential exam ordering, ignoring the interleaving of questioning, noisy observations, and exam cost.

Approach

A POMDP environment unifying patient questioning, exam ordering as tool calls, and diagnosis, with injected patient/exam noise and cost (CMS fees plus invasiveness tiers) in the reward; a small 8B agent is trained with SFT on synthetic Calgary-Cambridge conversations and then DAPO against a composite reward.

Key results

• On the OOD AgentClinic-MedQA set, MedExAgent-8B (Jac 0.378, Acc 0.378) is statistically tied with Aloe-Beta-70B (Jac 0.395, Acc 0.395) and outperforms the remaining baselines on Jac.
• On in-distribution DDxPlus, MedExAgent-8B reaches Jac 0.937 and Acc 0.939 versus 0.503 and 0.670 for the best baseline Aloe-Beta-70B.
• Removing the cost penalty raises average tool calls on DDxPlus from 3.75 to 5.65 and degrades DDxPlus Jac from 0.937 to 0.831.

Surprising findings

• Removing noise injection from training leaves DDxPlus performance essentially unchanged while it lowers PMC-Patients-v2 Jac by 0.088.
• Without the tool-match reward the model learns to call far fewer exams (2.38 vs 3.75 on DDxPlus), which hurts accuracy on AgentClinic.

Limitations

• Assumes the disease and observation function are static over the episode
• Training and evaluation rely on LLM-simulated patients rather than real patients
• English-only model

Experimental setup

Datasets: DDxPlus PMC-Patients-v2 AgentClinic-MedQA

Benchmarks: DDxPlus PMC-Patients-v2 AgentClinic-MedQA

Metrics: Sim (embedding cosine similarity) Jac (LLM-judge Jaccard) Acc (strict accuracy) Call F1 cost F1

Baselines: Qwen3-8B Qwen3-14B Qwen3-32B Llama-3.1-8B-Instruct Aloe-Beta-70B MedGemma-27B-text-it HuatuoGPT-o1-70B

Builds on: DAPO ToolRL Calgary-Cambridge model PatientSim Meditron3-8B

Closest papers by meaning

• Simulating Viva Voce Examinations to Evaluate Clinical Reasoning in Large Language Models (NeurIPS 2025 · similarity 0.89)
• MDAgents: An Adaptive Collaboration of LLMs for Medical Decision-Making (NeurIPS 2024 · similarity 0.88)
• MedChain: Bridging the Gap Between LLM Agents and Clinical Practice with Interactive Sequence (NeurIPS 2025 · similarity 0.88)
• MedAgent-Pro: Towards Evidence-based Multi-modal Medical Diagnosis via Reasoning Agentic Workflow (ICLR 2026 · similarity 0.88)
• MedAgentBoard: Benchmarking Multi-Agent Collaboration with Conventional Methods for Diverse Medical Tasks (NeurIPS 2025 · similarity 0.88)
• Doctor-R1: Mastering Clinical Inquiry with Experiential Agentic Reinforcement Learning (ICLR 2026 · similarity 0.88)