A recent study reveals that large language models (LLMs) tend to be overconfident in their responses, particularly on difficult tasks, while showing underconfidence on easier ones. The researchers introduced LifeEval, a new tool to assess model calibration based on task difficulty.