Simulation-trained language model agent more than halves under-triage of acute risks in psychiatric intake tests

The arXiv preprint reports under-triage fell from 30.5% to 12.9% across 256 adversarial simulated presentations, and psychiatrist raters preferred the RL-trained agent in 76.7% of 60 matched pairs played by psychiatrists, not real patients; the authors say gaps remain and require clinician oversight.