GPT-4.1-nano hits 93.5% accuracy, shows metacognitive sensitivity in medical reasoning
In 135 trials, GPT-4.1-nano's confidence tracked evidence quality—a rare calibration win for medical AI.
A new arXiv preprint by Ahmad Nazzal puts GPT-4.1-nano under a psychophysics-inspired microscope to ask a critical question: do large language models calibrate their confidence to the actual reliability of their medical reasoning? The study, titled "Large Language Models Show Metacognitive Sensitivity in Medical Reasoning," builds a controlled benchmark of 45 synthetic patient vignettes that weigh probable Alzheimer-type neurocognitive disorder (AT-NCD) against depression-related cognitive impairment (DRCI). Variations in evidence strength, conflicting clues, and missing information produced 135 forced-choice trials. The results: 93.5% diagnostic accuracy, 78.4% mean confidence, and an AUROC2 of 0.876—meaning confidence tracked correctness better than chance.
Notably, confidence rose as evidence moved away from the diagnostic boundary and fell when information was absent, and the model remained more confident on correct than incorrect trials even after adjusting for evidence strength and prompt format. That indicates partial metacognitive sensitivity, not globally uninformative confidence. But the smooth picture fractures in moderate, conflicting AT-NCD cases: GPT-4.1-nano shifted toward DRCI while retaining more confidence than accuracy justified—a localized calibration failure. The paper argues that confidence quality should be measured directly, not inferred from accuracy or model capability, and offers a reproducible framework for evaluating evidence sensitivity and calibration in medical LLMs.
- GPT-4.1-nano achieved 93.5% diagnostic accuracy across 135 medical trials with a mean confidence of 78.4%
- Confidence correctly decreased with missing information and tracked evidence distance, showing partial metacognitive sensitivity
- Errors clustered in moderate, conflicting Alzheimer's cases where the model overestimated its confidence, exposing calibration gaps
Why It Matters
For medical AI deployment, calibrated confidence is as critical as raw accuracy—this framework identifies where LLMs overtrust themselves.