Trained overnight model says 'ask a doctor' when unsure: errors drop from 46 to 13
MaziyarPanahi · x · 2026-10-28
The author trained a small open model overnight with a confidence-threshold strategy: it only answers on its own when 90%+ confident and hands everything else to a human with "not sure, ask a doctor."
Result: on cases it answered alone, mistakes dropped from 46 to 13. A simple, replicable selective-prediction / human-handoff pattern for high-stakes domains like healthcare.
Related event: Small open-source medical model learns to defer to doctors when uncertain(2 posts)→
More from Research
- Hobbyists find the "meth recipe neuron" in Talkie 1930 via single-neuron activation — Bedrovelsen · 2026-10-29
- ImagiRep paper: accurate imagined RGB isn't enough for spatial reasoning, GEARS picks better views — _akhaliq · 2026-10-29
- OpenAI paper claims to refute Exponential Time Hypothesis with subexponential 3-SAT algorithm — burny_tech · 2026-10-29
- Fireworks adds 31 partners to its practitioner-built Specialized Intelligence Index for real-work model evals — sophiamyang · 2026-10-29
- Will Anthropic's AI find cures? Scientist says the bottleneck is funding, not science — anshulkundaje · 2026-10-29
- NeurIPS paper: Discrete Neural Interlingua enables ontology-agnostic EHR modeling across hospitals — AI4Pathology · 2026-10-29