Google 用模拟精神科病人训练 LLM，鉴别诊断准确率从 33.3% 升至 68.3%

"Training Language Models in Simulation for Psychiatric Intake"

LLMs can answer medical questions, but conducting psychiatric interviews requires asking the right follow-up questions, recognizing safety risks, and communicating effectively with patients.

This Google paper trains an LLM using multi-turn reinforcement learning with simulated psychiatric patients, rewarding accurate diagnosis, effective communication, and clinical safety.

In blinded evaluations involving psychiatrists role-playing patients, the trained model outperforms its baseline in 76.7% of scenarios. Appropriate differential diagnoses improve from 33.3% to 68.3%, while patient-centered communication scores rise from 65.2% to 82.0%.

https://alphaxiv.org/abs/2610.40476