DeepMind RL-trained Gemini doubles correct psychiatric differential diagnoses in blind study
SRSchmidgall · x · 2026-11-04
Google DeepMind researchers (SRSchmidgall et al.) used multi-turn RL in simulation to train Gemini to conduct structured psychiatric intake interviews.
- Double-blind test: 32 psychiatrists role-played patients across 60 matched cases, blind-rated by an independent panel.
- The trained model beat an identically prompted base model in 76.7% of cases; appropriate differential diagnoses roughly doubled from 33.3% to 68.3%.
- Safety cuts both ways: under-triage of acute risk in adversarial scenarios dropped from 30.5% to 12.9% — a big improvement, but still roughly 1 in 8 missed.
The work shows simulated interactions alone can markedly improve clinical interviewing, though residual safety risk rules out direct deployment.
More from Models
- OCR on 2,165 pages of historic books: runaway loops down to 1, 1881 volume errors cut 38% — vanstriendaniel · 2026-11-04
- Median human turn-taking pause is -120ms, so voice AI's 500ms brag still misses how people talk — alexisgallagher · 2026-11-04
- Cloudflare Launches Clairic, a System One-Style Extraction Model Faster and Cheaper Than LLMs — michellechen · 2026-11-04
- Why a chatbot spells 17 as "seveen": tokenization quirks exposed when reasoning is off — mmitchell_ai · 2026-11-04
- TheZvi diffs OpenAI's new Model Spec: philosophically similar, but new examples dodge hard cases — TheZvi · 2026-11-04
- LlamaIndex stacks 17 small optimizations for ~10x faster OCRia-1 token generation on a single B200 — llama_index · 2026-11-04