#neuroscience

LLMs use hidden confidence signals to detect and fix their own errors

Research shows large language models maintain a second-order evaluative signal that predicts error detection and self-correction beyond what their output probabilities reveal.