基于隐藏层时序信号的 LLM 幻觉检测：快速傱里叶变换方法
计算与语言
2025-09-17 v1
摘要
幻觉仍是大型语言模型（LLM）在可靠性相关应用中面临的关键障碍。现有检测方法大多分为两类：事实性检查——受限于外部知识覆盖；以及静态隐藏状态分析，后者未能捕捉推理动态的偏差。因此，其有效性和鲁棒性仍有限。我们提出了HSAD（基于隐藏信号分析的检测），这是一种新颖的幻觉检测框架，通过建模自回归生成过程中隐藏表示的时序动态。HSAD 通过在不同层级上采样激活值构建隐藏层信号，应用快速傱里叶变换（FFT）获取频域表示，并提取最强非直流（non-DC）频分量作为特征。 Furthermore, by leveraging the autoregressive nature of LLMs, HSAD identifies optimal observation points for effective and reliable detection. Across multiple benchmarks, including TruthfulQA, HSAD achieves over 10 percentage points improvement compared to prior state-of-the-art methods. By integrating reasoning-process modeling with frequency-domain analysis, HSAD establishes a new paradigm for robust hallucination detection in LLMs.
引用
@article{arxiv.2509.13154,
title = {LLM Hallucination Detection: A Fast Fourier Transform Method Based on Hidden Layer Temporal Signals},
author = {Jinxin Li and Gang Tu and ShengYu Cheng and Junjie Hu and Jinting Wang and Rui Chen and Zhilong Zhou and Dongbo Shan},
journal= {arXiv preprint arXiv:2509.13154},
year = {2025}
}