arXiv 2602.147782026-05-19cs.CLcs.AIcs.CY

小尺寸语言模型幻觉的几何分析

A Geometric Analysis of Small-sized Language Model Hallucinations

Engineering (CEMSE) division, King Abdullah University of Science and Technology (KAUST)（卡塔尔科技大学工程学院（CEMSE））；Istituto di Informatica e Telematica (IIT), National Research Council of Italy (CNR)（意大利国家研究理事会信息与电信研究所（IIT））；Department of Information Engineering, University of Pisa（比萨大学信息工程系）

Emanuele Ricco, Elia Onofri, Lorenzo Cima, Stefano Cresci, Roberto Di Pietro

备注：: 30 pages, 12 figures, 14 tables, accepted as regular paper at ICML'26

阅读 PDF

arXiv 页面 ↗HTML 原文 ↗

AI总结：

本文从几何角度分析小尺寸语言模型幻觉问题，提出APORIA框架，通过句子嵌入空间研究重复提示下的响应，发现真实响应比幻觉响应更紧密聚类，并通过APORIA-LP方法实现高效分类，同时发布SOCRATES-300K数据集以支持进一步研究。

摘要：

幻觉--合理但事实错误的响应--对大型语言模型（LLMs）的可靠性构成重大挑战，尤其是在多步骤或代理设置中。现有研究大多将幻觉视为知识缺失的结果；我们显示，即使相关事实知识存在，模型仍会产生幻觉响应，这指向检索不稳定而非知识缺口。基于这一观察，我们引入APORIA（Aggregate Prompt-wise Observation Retrieving Instability via Asymmetry--幻觉所体现的困惑-矛盾状态），一种几何框架，研究相同提示的重复响应在句子嵌入空间中的情况。我们的核心假设是真实响应比幻觉响应更紧密聚类；我们实证验证了这一点，并显示经过Fisher投影后，两种响应类别变得一致可分离。我们通过APORIA-LP方法利用几何中的不对称性，这是一种高效的标签传播方法，通过最少30-50次注释即可对大量响应进行分类，在十种小尺寸LLM上实现超过90%的F1分数。为支持进一步研究，我们发布了SOCRATES-300K数据集，包含300,000个完全标注的响应，以及数据集生成和结果复现的代码。我们的关键发现--从嵌入空间的几何角度分析幻觉--补充了传统知识中心和单响应评估范式，为进一步研究铺平道路。

英文摘要：

Hallucinations -- plausible but factually incorrect responses -- pose a major challenge to the reliability of Large Language Models (LLMs), especially in multi-step or agentic settings. Existing work largely frames hallucinations as a consequence of missing knowledge; we show instead that, even when the relevant factual knowledge is present, models still produce hallucinated answers, pointing to retrieval instability rather than knowledge gaps. Building on this observation, we introduce APORIA (Aggregate Prompt-wise Observation Retrieving Instability via Asymmetry -- the state of puzzlement-in-contradiction that hallucinations embody), a geometric framework that studies repeated responses to the same prompt in sentence-embedding space. Our central hypothesis is that genuine responses cluster more tightly than hallucinated ones; we empirically validate this and show that, after Fisher projection, the two response classes become consistently separable. We leverage this asymmetry in geometry via APORIA-LP, an efficient label-propagation method that classifies large collections of responses from as few as 30--50 annotations, achieving F1 scores above 90% across ten small-sized LLMs. To support further research, we release SOCRATES-300K, a fully labelled dataset of 300,000 responses, together with the code for both dataset generation and result reproduction. Our key finding -- framing hallucinations from a geometric perspective in the embedding space -- complements traditional knowledge-centric and single-response evaluation paradigms, paving the way for further research.

机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。