MedExAgent: Training LLM Agents to Ask, Examine, and Diagnose in Noisy Clinical Environments
Researchers introduce MedExAgent, an AI system trained to perform clinical diagnosis through a POMDP framework that simulates real-world complexity including patient interaction, medical exams, and noisy data. The model uses supervised finetuning and reinforcement learning to balance diagnostic accuracy with cost-efficiency, achieving performance comparable to larger models while maintaining practical clinical constraints.
MedExAgent represents a meaningful advancement in medical AI by addressing a critical gap between simplified benchmarks and real-world clinical practice. Traditional medical LLM evaluations reduce diagnosis to single-turn questions or noise-free conversations, ignoring the interactive, iterative nature of actual healthcare. This research formalizes diagnosis as a POMDP with three action types—questioning, ordering exams, and issuing diagnoses—while introducing a systematic noise model with seven patient noise types and three exam noise types. This approach better reflects clinical reality where doctors must navigate incomplete information, patient variability, and resource constraints.
The training methodology combines supervised finetuning on synthetic conversations structured after the Calgary-Cambridge clinical interview model with DAPO (a reinforcement learning technique) that optimizes a composite reward function. This reward balances diagnostic accuracy, tool-use quality, and exam costs including both financial and patient comfort factors. The approach demonstrates that smaller, specialized models can match larger general-purpose LLMs on diagnostic tasks while maintaining efficiency—a crucial consideration for healthcare deployment.
For the medical AI industry, MedExAgent's framework offers a more rigorous evaluation standard that could reshape how diagnostic systems are benchmarked and deployed. Healthcare systems face mounting pressures to reduce costs while improving outcomes, making cost-efficient AI diagnostics particularly valuable. The emphasis on handling noisy, incomplete data and patient variability makes this work immediately relevant for real-world implementation rather than academic performance metrics.
Future development should focus on clinical validation with actual patient data, integration with existing electronic health records systems, and evaluation across diverse patient populations and disease categories to ensure robustness and reduce algorithmic bias.
- →MedExAgent trains agents to perform interactive diagnosis by asking questions, ordering exams, and managing costs—more accurately modeling real clinical workflows than existing benchmarks
- →The two-stage training pipeline combines supervised finetuning on structured interview data with reinforcement learning to balance diagnostic accuracy, exam quality, and cost efficiency
- →A systematic noise model with ten noise types enables training on realistic clinical scenarios with incomplete and variable patient information
- →The approach achieves diagnostic performance comparable to larger models while maintaining lower examination costs and resource utilization
- →The POMDP framework and reward optimization method provide a reusable benchmark and methodology for evaluating diagnostic AI systems in uncertainty-rich environments