Psychiatric assessment starts with a conversation: clinicians must build trust while asking questions that uncover symptoms, risks, and history. Mistakes in this intake can lead to wrong diagnoses or missed safety concerns. Large language models, software that generates human-like text, can converse, but they are usually trained on written text, not on conducting interviews.
The researchers trained such a model using , a method where the model practices in simulated patient encounters and is rewarded for accurate questioning, empathic responses, and safe decisions. Its performance was compared with an identical model that skipped this practice. Dozens of board-certified psychiatrists played patients in text-based encounters, while independent psychiatrists scored the conversations without knowing which model was which.
The trained model won in 76.7% of matched scenario pairs and showed stronger patient-centered communication and diagnosis. It also under-triaged serious safety risks about half as often. Simulation-based practice appears to transfer to human conversations, though residual errors mean human oversight remains necessary.