Google trains Gemini with RL on simulated patients; double-blind study shows big psychiatric intake gains
SRSchmidgall · x · 2026-10-31
- Google DeepMind/Research/Health teams published an arXiv paper (2610.40476) training Gemini via multi-turn RL in simulated-patient environments to conduct structured psychiatric intake: history-taking, differential diagnosis, and initial management planning
- Method: multi-turn RL in simulation covering the full clinical interview workflow
- Double-blind crossover study: 60 matched scenario pairs, 120 encounters, 32 psychiatrists role-playing patients, comparing the RL-trained agent against its identically prompted baseline
- Result: RL in simulation significantly improves Gemini's performance on structured text-based psychiatric intake
- Led by first author Samuel Schmidgall with 55 co-authors spanning hospitals and research institutions
Related event: Google trains Gemini for psychiatric intake via simulated-patient RL(11 posts)→
More from Models
- OpenBMB open-sources MiniCPM-OCR: 1.3B model parses 128 pages per minute — haider1 · 2026-10-31
- MiniCPM-OCR: 1.3B model does 128 pages/min, 2B checker nearly matches 397B model — haider1 · 2026-10-31
- Anthropic renames Max plans, drops 5-hour limit for verified Ultra accounts — rohanpaul_ai · 2026-10-31
- Anthropic's Ultra tier drops 5-hour rate limits, confirms Alex Albert on X — EricBuess · 2026-10-31
- Claude refuses three times to add logos to a slide while ChatGPT just says 'Done' — flavioAd · 2026-10-31
- Gemini App says it was optimized for the new iPhone Duo on day one — GeminiApp · 2026-10-31