Google DeepMind trains Gemini 3.1 Flash-Lite in simulation for psychiatric intake

A team from Google DeepMind and Google Research has posted a preprint, Training language models in simulation for psychiatric intake, submitted on 29 October. They took a small model, Gemini 3.1 Flash-Lite, and trained it with multi-turn reinforcement learning against simulated patients, then had psychiatrists judge whether it got better at the first structured interview a psychiatrist does with a new patient.

How it was tested

The simulated patients are built so that key facts only come out if the agent builds rapport and asks the right questions. Per the full text, training follows DeepMind’s earlier ResidencyRL framework.

The main human test was a randomized, double-blind crossover: 32 board-certified psychiatrists played standardized patients, talking with both the trained agent and the identically prompted base model, and a separate panel of 15 psychiatrists scored the de-identified transcripts.

Results the paper reports

• Preference: the trained agent won in 76.7% of 60 matched scenario pairs.
• Differential diagnosis judged appropriate: 68.3% versus 33.3% for the base model.
• Simulated scenarios: across 127 psychiatrist-validated cases, the clinical rubric score rose from 0.379 to 0.706, as graded by a Gemini 3.1 Pro autorater.
• Safety: on 256 adversarial presentations, under-triage of acute safety risks fell from 30.5% to 12.9%.

The limits the authors state

Psychiatrists played the patients specifically to avoid exposing real patients to an unvalidated system. The work covers single-encounter, text-only intake, not ongoing care. The ground-truth diagnoses and treatment recommendations for the 48,243 training scenarios were generated by a language model and not individually reviewed by clinicians. The authors say residual gaps “require safeguards and clinician oversight”. Nothing in the paper describes a product or clinical use.

One detail deserves attention beyond medicine. The headline rubric gain was graded by a Gemini model, the same family the agent was trained against. When the authors re-graded with GPT-5.6-Luna as an independent judge, the gain held but shrank, from a 0.270 gap to 0.165.

Google’s other recent health result, the AMIE pre-visit study, did involve real patients at one clinic; this preprint stays in simulation and role-play.

Why this matters

For anyone training agents, this is a clear case of multi-turn RL in a simulator lifting a cheap model on a long conversation, with a human-graded check rather than only automated scores. The judge result is the practical lesson: an independent judge measured a gain about 40 percent smaller than the same-family judge did, so cross-check with a judge from another vendor, as in our guide to evaluations for small teams. The clinical claims stay at preprint level.