DeepMind: RL on simulated patients cuts AI psychiatry missed risk escalation from 30.5% to 12.9%
rohanpaul_ai · x · 2026-10-31
A Google DeepMind paper shows multi-turn RL against simulated psychiatric patients trained Gemini 3.1 Flash-Lite into a better intake interviewer. In a double-blind study with 32 psychiatrists role-playing patients and 15 blind graders, the trained agent beat its base model in 76.7% of 60 matched pairs; missed escalations of acute risk on 256 adversarial cases dropped from 30.5% to 12.9%, and differential diagnosis appropriateness rose to 68.3% vs 33.3%. Gaps remain — only 46.7% of encounters got comprehensive treatment plans — so clinicians must stay in the loop.
More from Models
- GPT-6 Luna, Terra, Sol and Astra: rumored OpenAI model names surface — Angaisb_ · 2026-11-01
- Xiaomi's OCR model Xiaomi-OCR-0 trends on Hugging Face — xiaomi-research · 2026-11-01
- DeepSeek-V4.1-Flash tops Papers with Code trending methods at 20.1x growth — NielsRogge · 2026-11-01
- GPT-6 Astra field guide: top in coding and science, #16 in finance, 21% freelance pass rate — Hrstar1 · 2026-11-01
- ThursdAI weekly: Gemini 4 Carbon takes #1, Anthropic files to go public — thursdai_pod · 2026-11-01
- Anthropic's usage restrictions spark backlash over paid-model rights and open-weight futures — QuixiAI · 2026-10-31