Google trains Gemini via RL on simulated patients, halving psychiatric intake safety failures

SRSchmidgall · x · 2026-10-31

Google DeepMind and academic partners (Harvard, Stanford, Johns Hopkins, etc.) published a paper training Gemini with reinforcement learning in simulated-patient scenarios for structured psychiatric pre-visit intake.

Key results

In a double-blind study with 32 psychiatrists role-playing patients, multi-turn simulation RL substantially improved diagnostic formulation, clinical skills (PACES), treatment planning, and clinical safety.
Acute safety under-triage on adversarial presentations dropped from 30.5% to 12.9% (p < 0.0001); contraindication errors fell from 6.2% to 0.8% (p = 0.001).
Substantial gaps remain: comprehensive, appropriate treatment was recommended in only 46.7% of OSCE encounters, and 1 in 8 adversarial acute-risk cases were still under-triaged.

Safety & evaluation methodology

Two multidisciplinary Delphi panels built consensus taxonomies (21 clinical safety risks across 6 domains; 18 relational competencies across 5 domains), yielding 256 adversarial scenarios.
Residual failures are mainly errors of omission (questions never asked) that post-hoc output filters cannot catch, requiring process-level safeguards external to free-text generation.
The simulation autorater agrees moderately with psychiatrists (ρ = 0.57–0.58): reliable for detecting the direction of model gains, not their magnitude.

The authors stress this is a strictly assistive, clinician-supervised research prototype, not a replacement for psychiatric care, and further work is needed before prospective clinical validation with real patients.

Original post →

Google trains Gemini via RL on simulated patients, halving psychiatric intake safety failures

More from Safety