提出证据图一致性检测方法，发现不同大模型的幻觉模式截然不同
Evidence Graph Consistency in Retrieval-Augmented Generation: A Model-Dependent Analysis of Hallucination Detection
- 构建每条回答的局部证据图，用5种结构一致性指标衡量幻觉
- 在6个大模型上测试，发现Llama-2与GPT系列幻觉信号方向相反
- 揭示幻觉检测需针对模型家族定制，不能通用嵌入式图结构方法
检索增强生成（RAG）虽能降低大语言模型的幻觉，但无法彻底消除。现有检测方法依赖生成答案与检索片段之间的简单相似性，忽略了证据间和答案主张之间的结构关系。本文提出证据图一致性（EGC）框架，为每条响应构建局部证据图，并计算五种结构一致性度量作为幻觉指示器。在RAGTruth数据集完整问答划分上对六个大模型（共5,767条响应）进行评估，发现模型族间的显著差异：在Llama-2模型中，图一致性特征呈现预期的幻觉诊断方向；但在GPT-4、GPT-3.5和Mistral-7B中则出现系统性反转。这一反转表明不同模型家族存在质异的幻觉模式，说明基于嵌入的图一致性无法作为普适的幻觉检测信号。
原文摘要 · Abstract (English)
Retrieval-Augmented Generation (RAG) reduces but does not eliminate hallucination in large language models. Existing detection methods rely on flat similarity between generated answers and retrieved passages, ignoring structural relationships among evidence pieces and answer claims. We propose Evidence Graph Consistency (EGC), a framework that constructs a local evidence graph per response and computes five structural consistency measures as hallucination indicators. Evaluated on the full question answering split of RAGTruth across six LLMs (5,767 responses), EGC reveals a consistent model-family split: graph consistency features show the expected diagnostic direction for hallucinations in Llama-2 models but exhibit systematic reversal in GPT-4, GPT-3.5, and Mistral-7B. This reversal suggests qualitatively different hallucination patterns across model families and indicates that embedding-based graph consistency cannot serve as a model-independent hallucination detection signal.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。