Google DeepMind 用模拟病人训练 Gemini 3.1 Flash-Lite 做精神科问诊
DeepMind 让模型在模拟病人身上练问诊，漏报急性风险从 30.5% 降到 12.9%，做医疗 Agent 的可以看看方法。
Google DeepMind 在 arXiv 论文（2610.40476）中用多轮强化学习训练 Gemini 3.1 Flash-Lite，与只在建立信任后才透露关键信息的模拟病人对话。32 名精神科医生扮演病人，另外 15 名盲评对话记录，训练后的模型在 60 组配对案例中 76.7% 胜过同提示词的基座模型。在 256 个对抗性案例中，急性风险漏报率从 30.5% 降到 12.9%。不过仍有约八分之一的对抗性危机未被升级上报，作者建议保留临床医生监督。
New Google DeepMind paper shows that training a model with RL on simulated patients made it a better psychiatric intake interviewer, as judged by psychiatrists.
training on simulated psychiatric intakes cut missed escalations of acute risk from 30.5% to 12.9% in adversarial cases.
In psychiatry, the intake interview is a main diagnostic tool, and key facts often surface only through trust and careful questions. Most conversational medical AI evaluations have left psychiatric cases out.
They trained Gemini 3.1 Flash-Lite with multi-turn RL against simulated patients who reveal key facts only after rapport and targeted questions. Then 32 psychiatrists role-played patients, while 15 others graded the transcripts blind.
The trained agent beat its identically prompted base model in 76.7% of 60 matched case pairs. On 256 adversarial cases, it failed to escalate acute risks 12.9% of the time, down from 30.5%.
If you build health agents, train them on simulated conversations, but keep clinicians supervising, since about 1 in 8 adversarial crises went unescalated.
– arxiv. org/abs/2610.40476
Title: "Training language models in simulation for psychiatric intake"
