Hallucination Detection in LLMs Using Spectral Features of Attention Maps

TL;DR

This paper introduces a spectral analysis method using Laplacian eigenvalues of attention maps to improve hallucination detection in Large Language Models, achieving state-of-the-art results.

Contribution

The paper proposes the $ ext{LapEigvals}$ method that leverages spectral features of attention maps for more effective hallucination detection in LLMs.

Findings

Achieves state-of-the-art performance among attention-based methods.

Demonstrates robustness and generalisation across different models.

Provides extensive ablation studies validating the approach.

Abstract

Large Language Models (LLMs) have demonstrated remarkable performance across various tasks but remain prone to hallucinations. Detecting hallucinations is essential for safety-critical applications, and recent methods leverage attention map properties to this end, though their effectiveness remains limited. In this work, we investigate the spectral features of attention maps by interpreting them as adjacency matrices of graph structures. We propose the $\text{LapEigvals}$ method, which utilises the top-$k$ eigenvalues of the Laplacian matrix derived from the attention maps as an input to hallucination detection probes. Empirical evaluations demonstrate that our approach achieves state-of-the-art hallucination detection performance among attention-based methods. Extensive ablation studies further highlight the robustness and generalisation of $\text{LapEigvals}$, paving the way for future advancements in the hallucination detection domain.