Google's ResidencyRL: AI Learns Clinical Skills Through 50K Simulated Patient Encounters

SRSchmidgall · x · 2026-08-12

Google researchers introduced ResidencyRL, a framework putting Gemini 3.5 Flash through a simulated medical residency.

Scale: The model trained across 49,870 simulated telehealth encounters covering 81 conditions.
Adversarial Environment: AI patients hid symptoms, resisted advice, and requested inappropriate treatments, forcing the model to gather info over conversations up to 60 turns.
Results: Diagnostic accuracy under adversarial conditions rose from 81% to 88%, while missed red flags fell by 31%.
Blind Evaluation: Clinicians preferred the trained agent over the base model in 87.6% of 97 comparisons.

The improvements also successfully transferred to unseen oncology cases and external benchmarks.

More from Models