Detecting Hallucinations in Large Language Models via Internal Attention Divergence Signals
This work offers a lightweight, interpretable uncertainty signal for LLM hallucination detection, but is incremental as it builds on known attention-based methods.
The authors propose a single-pass method using attention divergence from a uniform distribution to detect hallucinations in LLMs, achieving competitive performance with existing uncertainty methods across multiple datasets and models.
We propose a lightweight and single-pass uncertainty quantification method for detecting hallucinations in Large Language Models. The method uses attention matrices to estimate uncertainty without requiring repeated sampling or external models. Specifically, we measure the Kullback-Leibler divergence between each attention head's distribution and a uniform reference distribution, and use these features in a logistic regression probe. Across multiple datasets, task types, and model families, attention divergence is highly predictive of answer correctness and performs competitively with existing uncertainty estimation methods. We find that this signal is concentrated in middle layers and on factual tokens such as named entities and numbers, suggesting that attention dynamics provides an efficient and interpretable white-box signal of model uncertainty.