潜在事实核查：通过激活工程检测虚假信息
原标题：Latent Fact-Checking: Detecting Misinformation through Activation Engineering
AI 摘要
该研究提出一种基于激活工程的虚假信息检测框架，通过对比真实与虚假陈述的激活差异提取“虚假方向”，并在推理时将未知声明的最后token激活投影到该方向进行分类，无需微调或外部知识检索。方法在Gemma、Llama和Qwen家族的11个模型（270M至12B参数）上评估，在LIAR和FACTors基准上匹配或超越零样本和少样本提示基线，尤其在小模型上提升显著，但在AVeriTeC上表现有限。研究证明预训练语言模型的潜在空间中真实性是线性可分的结构概念。
正文节选
Computer Science > Machine Learning Title:Latent Fact-Checking: Detecting Misinformation through Activation Engineering View PDF HTML (experimental) Abstract:The proliferation of misinformation online has driven demand for scalable detection systems. While most existing approaches rely on surface-level linguistic features or external knowledge retrieval, we examine truthfulness as a geometric property of a language model's representation space. We introduce a misinformation detection