Understand this story
Google DeepMind published a paper introducing multi-turn reinforcement learning training of Gemini 3.1 Flash-Lite using simulated patients to improve performance in psychiatric intake interviews. According to reports on the paper, in 60 paired cases, 76.7% were judged by psychiatrists in blinded evaluations to be superior to the base model with the same prompt; in 256 adversarial cases, the rate of acute risk not being escalated fell from 30.5% to 12.9%, meaning that roughly 1 in every 8 cases still had a crisis that was not reported. In training and evaluation, 32 psychiatrists played the role of patients, and another 15 doctors blindly evaluated the dialogue records. Currently, this paper is the only source of the report, and there have been no subsequent developments.
AI-generated from public reports
Latest development · Report
DeepMind trains Gemini psychiatric consultations with simulated-patient RLDevelopments
1 confirmed developments
Reports and discussion timeline
Browse reports and discussion by time. Their evidence levels remain distinct.
Oct 31
Rohan Paul
DeepMind trains Gemini psychiatric consultations with simulated-patient RL
Google DeepMind used simulated patients to conduct multi-turn RL training on Gemini 3.1 Flash-Lite; in 60 paired cases, 76.7% were judged by psychiatrists in a blinded evaluation to be superior to the base model with the same prompt. In 256 adversarial cases, the rate of acute risk not being escalated fell from 30.5% to 12.9%, meaning that roughly 1 in every 8 cases still had a crisis that was not reported. 32 psychiatrists played the role of patients, and another 15 doctors blindly evaluated the conversation records.