Simulated-patient training reportedly improves Gemini’s psychiatric intake performance

Researchers report that Gemini, trained through multi-turn reinforcement learning with simulated patients, outperformed its base model in text-based psychiatric intake encounters. In a study with psychiatrists portraying patients, the trained model scored higher on patient-centered communication (82.0% vs. 65.2%) and differential diagnosis appropriateness (68.3% vs. 33.3%). The paper also reports lower acute-risk under-triage in separate adversarial simulations. The authors say remaining gaps require safeguards and clinician oversight; the study did not involve actual patients.