health
LLM Confidence Was Often Overstated on Neurology Questions
A medRxiv preprint tested eight general-purpose large language models on 100 neurology board-style questions. Confidence estimates were often too high, while agreement across repeated answers best distinguished correct from incorrect responses.