This study tackles token-level hallucination detection in outputs of large language models. Previous studies revealed that attention exhibits irregular patterns when hallucination occurs. Inspired by this, we extract features from the attention matrix that provide complementary views of (a) the average attention each token receives, which helps identify whether certain tokens are overly influential or ignored, (b) the diversity of attention each token receives, which reveals whether attention is biased toward specific subsets, and (c) the diversity of tokens a token attends to during generation, which indicates whether the model references a narrow or broad range of information. These features are input to a Transformer-based classifier to conduct token-level classification to identify hallucinated spans. Experimental results indicate that the proposed method outperforms strong baselines on hallucination detection with longer input contexts, i.e., data-to-text and summarization tasks.
View on arXiv@article{ogasa2025_2504.04335,
title={ Hallucination Detection using Multi-View Attention Features },
author={ Yuya Ogasa and Yuki Arase },
journal={arXiv preprint arXiv:2504.04335},
year={ 2025 }
}