Reinforcement learning cut missed acute-risk escalations from 30.5% to 12.9% in adversarial safety tests, according to a new preprint
Google DeepMind researchers have put a small Gemini model through something like a psychiatry residency. In a preprint published October 29, 2026, they report that training Gemini 3.1 Flash-Lite against simulated patients meaningfully improved how it conducts psychiatric intake interviews.
The headline safety number is the one that matters most. In adversarial tests, the share of acute safety risks the model failed to escalate fell from 30.5% to 12.9%.
What the researchers actually tested
The paper, titled “Training language models in simulation for psychiatric intake” (arXiv:2610.40476), applies multi-turn reinforcement learning to Gemini 3.1 Flash-Lite. The model talks with simulated patients, gets graded, and adjusts.
To judge the results, the team ran a randomized, double-blind crossover objective structured clinical examination, or OSCE. Thirty-two board-certified psychiatrists compared the reinforcement-learning-trained agent against the baseline model without knowing which was which.
The trained agent won 76.7% of those paired comparisons, with a reported p-value below 0.0001.
The gains showed up in several areas the psychiatrists scored:
- Alliance and communication: 82.0% for the trained agent versus 65.2% for the base model
- Diagnostic appropriateness: 68.3% versus 33.3%
- Improvements in history-taking, differential diagnosis, management and documentation
Across 127 psychiatrist-validated scenarios, the trained agent’s clinical rubric score rose to 0.706 from 0.379.
The safety test that matters most
The under-triage figure comes from 256 adversarial scenarios built around standardized safety-risk classifications. The baseline model missed acute-risk escalations 30.5% of the time. After training, that dropped to 12.9%, again with a p-value under 0.0001.
Building on ResidencyRL
The work extends DeepMind’s previously established ResidencyRL framework into psychiatry. A central design choice is a hierarchical reward system that ranks priorities, putting clinical competencies alongside safety metrics.
The research also uses a deliberately lightweight model. Flash-Lite sits at the smaller, cheaper end of the Gemini lineup.
What this means for healthcare AI
There are clear limits worth watching. This is a preprint, so the findings have not yet completed formal peer review, and all of the testing happened in simulated and examination settings rather than with real patients.
A drop to 12.9% under-triage in adversarial scenarios is a strong research result, yet any deployment in consequential care would presumably need human oversight and much more validation.
Thirty-two psychiatrists, blinded to which model they were grading, consistently picked the trained one.
