Psychiatric maintenance: trained in simulation, a language model prevails in 77% of cases against 32 psychiatrists playing patients

Simulated environmental reinforcement learning doubles the clinical score of the model and reduces the underestimated suicidal risks by more than half.

A team led by Samuel Schmidgall led a language model to conduct psychiatry admission interviews (history collection, differential diagnosis, first management decisions) through strengthening learning on multi-tour conversations in simulated environments.

Double-blind assessment

To test it, 32 graduate psychiatrists played the role of standardized patients against the trained agent and basic model, on 60 pairs of matched scenarios (120 consultations). A second panel of 15 independent psychiatrists noted anonymized transcripts. The trained agent was preferred in 76.7% of the pairs; This included 68.3% of differential diagnoses considered appropriate compared to 33.3%, and 82% of patient-centred diagnoses compared to 65.2%. Of 127 simulated scenarios validated by psychiatrists, his clinical score almost doubled (0,706 compared to 0.379).

The authors also constructed 256 difficult safety situations by Delphi consensus: underestimation of acute risks fell from 30.5 per cent to 12.9 per cent. They stress that these residual deviations require safeguards and clinical supervision.