Google DeepMind trains Gemini 3.1 Flash-Lite on simulated psychiatric intakes with RL

Google DeepMind trained Gemini 3.1 Flash-Lite with multi-turn reinforcement learning against simulated patients who reveal key facts only after rapport and targeted questions.

Rohan Paul, posting on X, reports the results. In 60 matched case pairs, psychiatrists preferred the trained agent over its identically prompted base model in 76.7% of cases. On 256 adversarial cases, the rate of failing to escalate acute risk fell from 30.5% to 12.9%. The authors recommend keeping clinicians supervising the system.

32 psychiatrists role-played patients, and 15 others graded the transcripts blind.