Medical AI first learns to say it doesn't know: a 4B model speaks only when 90% confident
Trained on a single A100 for 27 minutes, at a cost of about $1.15. The model answers on its own only when more than 90% confident; otherwise it hands off to doctors. On 1,273 unseen exam questions, accuracy rose from 66.7% to 74.5%. Scores went up, but the more interesting part is when it keeps quiet.
Maziyar Panahi posted a very small medical model experiment on X. The base model is Qwen3.5-4B, combined with UnslothAI's decision fine-tune; it ran for 27 minutes on a single A100, with a bill of about $1.15.
A hard rule was added to the model: it answers on its own only when confidence is above 90%, and everything else is handed off to humans. Maziyar says that when the model answered by itself, errors dropped from 46 to 13. On 1,273 unseen MedQA exam questions, accuracy rose from 66.7% to 74.5%.
The video shows a medical decision question. An orthopedic resident has completed carpal tunnel repair and discovers that a flexor tendon was cut by mistake; what should be done next? Option A is to disclose the error to the patient, shown at 76%; B is to tell the attending physician, 19%; C is to report the doctor to a superior, 1%; D is to refuse to document the surgical procedure, 4%. The screen also notes the rule: the model decides on its own only when confidence is above 90%; otherwise it hands off to a clinician.
The training data is MedQA-USMLE-4-options, CC BY 4.0, hosted on HuggingFace; the dataset was introduced by Jin et al. in a 2020 paper. The link is here: MedQA-USMLE-4-options. It should be clear that these are exam questions, with no patient data, and are not medical advice.
The most interesting part of this experiment is not the 74.5%. A 4B model getting 8 percentage points more exam questions right is not explosive by today's model leaderboard standards. What is genuinely interesting is the 90% threshold: it splits answering into two actions—answer if it can, hand off to a human if it cannot. When the model answered by itself, errors fell from 46 to 13, showing that part of why the system became more accurate is that it began to admit what it does not know.
In the comments, Rajan G mentions that putting a reasoning model in front of a decision model and adding a threshold for saying I don't know is crucial in clinical and clinical-research workflows. LukeTestsAI offers a rougher rule: it is fine to let the bot run errands—reading things, sorting, tracking packages—but anything involving spending money or messaging a real person still has to wait for his yes.
Both comments point to the same issue. The hard part of medical AI is not just getting the question right; it is also stopping when uncertain and having someone to catch the process after it stops. Training a small model that knows how to refuse may cost only $1.15. The people, processes, and accountability needed to catch those 13 errors are not cheap.
Maziyar ends by asking which decisions you would be willing to let AI make on its own. That question is harder to answer than the model's score.
发布时间: 2026-10-29 00:38