Detecting Hallucinations in Large Language Models via Internal Attention Divergence Signals

Gijs van Dijk

We propose a lightweight and single-pass uncertainty quantification method for detecting hallucinations in Large Language Models. The method uses attention matrices to estimate unc…