ResidencyRL preprint reports improved clinical AI decisions in simulation

Researchers report in a preprint that ResidencyRL, a method for training clinical AI agents through simulated patient encounters, improved diagnostic accuracy under adversarial conditions to 88.0% from 81.0% and reduced missed red flag rates by 31%. The method uses reinforcement learning to reward diagnostic accuracy, management quality, communication, documentation and safety across encounters lasting up to 60 dialogue turns and 8 tool calls. Language models simulate patients with complex, adversarial behaviors. Blinded expert clinicians preferred the trained agent in 87.6% of side-by-side comparisons. The agent also outperformed its base model across all six clinical axes of the AMIE multi-visit benchmark. Prospective validation in real-world workflows remains necessary to establish clinical utility.