# Research Team Trains Gemini 3.1 Flash-Lite for Psychiatric Intake

By Simon Yoon

Canonical URL: https://www.tokenpost.com/news/technology/45554
Published: 2026-10-31T15:37:55.000Z
Updated: 2026-10-31T15:37:55.000Z
Section: Technology

> Psychiatrists rated it higher in 46 of 60 matched pairs, but it failed to recommend urgent or emergency care in 33 of 256 safety scenarios.

A research team trained Gemini 3.1 Flash-Lite to conduct psychiatric intake in simulated interviews, earning higher psychiatrist ratings in most matched comparisons while still missing some acute-care escalations.

The team included researchers affiliated with Google DeepMind, Google Research, Google for Health and academic institutions. It used multi-turn reinforcement learning and simulated patient encounters to train the model.

A blinded review covered 120 text consultations arranged as 60 matched scenario pairs. Thirty-two board-certified psychiatrists role-played standardized patients. A separate panel of 15 psychiatrists rated de-identified transcripts without knowing which model produced them. The trained model received higher aggregate ratings in 46 of 60 matched pairs, or 76.7%. Both versions used identical prompts and tools.

On an adapted clinical-skills rubric scored from one to five, the trained model averaged 3.04, compared with 2.32 for the base model. The trained model’s median consultation lasted 60 clinician turns, versus eight for the base model. The encounters had no turn limit.

A separate set of 256 adversarial safety scenarios tested acute psychiatric crises. Under-triage, or missing an active, life-threatening crisis requiring urgent or emergency care, fell to 12.9% for the trained model from 30.5% for the base model. The trained model still failed to recommend urgent or emergency psychiatric care in 33 of 256 scenarios.

In another evaluation, Gemini 3.1 Pro scored the trained model on 127 psychiatrist-validated simulated cases. It assigned the trained model a clinical-rubric score of 0.706, compared with 0.379 for the base model. This was an automated assessment of simulated cases, not a measure of outcomes in real patients.

The research relied entirely on simulated cases rather than real patient records. Prospective testing with real patients and clinician oversight remains a next step. The study’s authors wrote, “Such text-based systems are not a substitute for psychiatric clinicians and are not ready for autonomous psychiatric assessment.”

Google announced Gemini 3.1 Flash-Lite for preview through the Gemini API and Vertex AI on March 3, 2026. The model is designed for fast, cost-efficient performance on high-volume, latency-sensitive tasks.