Google Trains Gemini for Psychiatric Intake via RL with Simulated Patients, Halving Missed Triage

Google DeepMind, together with Harvard, Stanford, Johns Hopkins and over a dozen other institutions, published the paper "Training language models in simulation for psychiatric intake" (arXiv:2610.40476, a 56-author team led by Samuel Schmidgall). The work trains Gemini via multi-turn reinforcement learning in a simulated-patient environment to conduct structured psychiatric intake interviews, focusing on two skills unique to psychiatry: history taking and building patient rapport. Simulation-based training markedly improved visit safety, with a large drop in safety-related missed calls.

Confirmed

In a double-blind OSCE (Objective Structured Clinical Examination) with 32 psychiatrists role-playing patients, the simulation-trained agent beat the base model in 76.7% of paired comparisons: clinical examination score 3.04 vs 2.32, differential diagnosis accuracy 68.3% vs 33.3%, communication ability 82.0% (base model figure not fully given in the post)
Acute under-triage rate fell from 30.5% to 12.9% (p<0.0001); contraindicated prescriptions dropped from 6.2% to 0.8% (p=0.001)
Across 127 simulated intake scenarios validated by 54 board-certified psychiatrists, clinical rubric score rose from 0.379 to 0.706 and composite score from 0.294 to 0.564, cross-checked by an independent external family review model (GPT-5.6 series)

Why it matters

Psychiatric intake involves high-risk triage decisions and contraindicated prescriptions, where missed cases and wrong prescriptions carry enormous safety costs. The work shows that reinforcement learning in simulated environments can substantially improve a model's clinical safety without real patients, offering a verifiable technical path for LLMs to enter highly sensitive medical settings.

2026-10-31 ~ 2026-10-31 · 12 related posts

Primary sources

Google trains Gemini via RL on simulated patients, halving psychiatric intake safety failures — SRSchmidgall · 2026-10-31 00:32 (UTC+08:00)
Google's 56-author paper: training LLMs in simulation for psychiatric intake — SRSchmidgall · 2026-10-31 00:32 (UTC+08:00)
Simulation RL cuts Gemini psychiatric under-triage from 30.5% to 12.9% — SRSchmidgall · 2026-10-31 00:32 (UTC+08:00)

[source] Google trains Gemini via RL on simulated patients, halving psychiatric intake safety failures — SRSchmidgall · 2026-10-31
Double-blind OSCE: simulation-trained agent wins 76.7% of psychiatrist pairs — SRSchmidgall · 2026-10-31
Simulation-trained Gemini nearly doubles clinical rubric score, 0.706 vs 0.379 — SRSchmidgall · 2026-10-31
Simulated eval reliably detects direction of model gains, not magnitude — SRSchmidgall · 2026-10-31
[source] Simulation RL cuts Gemini psychiatric under-triage from 30.5% to 12.9% — SRSchmidgall · 2026-10-31
[source] Google's 56-author paper: training LLMs in simulation for psychiatric intake — SRSchmidgall · 2026-10-31
DeepMind's ResidencyRL: simulated RL lifts psychiatric intake AI, preferred in 76.7% of cases — SRSchmidgall · 2026-10-31

5 near-duplicate retellings: SRSchmidgall · SRSchmidgall · SRSchmidgall · SRSchmidgall · SRSchmidgall