跳转至

大语言模型中的知识归因探测研究

Probing for Knowledge Attribution in Large Language Models

arXiv:2602.22787 [cs.CL]
Subjects: Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
Authors: Ivo Brink, Alexander Boer, Dennis Ulmer
Submitted: 26 Feb 2026; Last revised: 10 Sep 2026 (v3)
DOI / Paper: 10.48550/arXiv.2602.22787

文章背景与核心概要

大语言模型 (Large Language Model, LLM) 在实际落地中频繁出现“幻觉”现象。这些错误通常源于两大根源:其一是误读或篡改了给定的提示词上下文(违背忠实性),其二则是模型内部参数化记忆出现了事实偏差(违背真实性)。要彻底缓解幻觉,核心在于厘清模型生成特定答案时究竟主导调用了哪一部分知识——即实现精准的“贡献性知识归因” (Contributive Attribution)。

本文指出,仅需在大模型中间隐藏层表征上训练轻量级的线性探测器 (Linear Probes) ,就能以极高置信度识别输出内容的知识来源。研究团队提出了 AttriWiki 自监督数据合成管线,摆脱了以往必须依赖人为制造知识冲突的缺陷,能够全自动生成高质量标签样本。基于 AttriWiki 训练的探测器在 Llama-3.1-8B、Mistral-7B 和 Qwen-7B 等模型上实现了高达 0.96 的 Macro-\(F_1\) 得分,并能无缝零样本迁移至 SQuAD 与 WebQuestions 等经典问答基准,为大模型的可信部署与可解释性分析开辟了全新路径。


概述

Summary

大语言模型经常受到幻觉问题的困扰,这些幻觉通常源自两个截然不同的诱因:忠实度违背 (Faithfulness Violations)(即误用或曲解了提示词所提供的上下文)以及 真实度违背 (Factuality Violations)(即根植于模型自身内部参数化知识的事实性错误)。要有效化解这些问题,精准定位模型输出背后究竟主要由哪个知识源头主导——这一过程被称为贡献性归因 (Contributive Attribution)——显得至关重要。

Large language models often suffer from hallucinations, which typically stem from two sources: faithfulness violations (misusing provided context) and factuality violations (errors originating from internal parametric knowledge). Mitigating these issues requires accurately identifying the dominant knowledge source behind an output—a process known as contributive attribution.

本研究表明,在模型隐藏层表征上训练的简单线性探测器 (Linear Probes) ,能够极其可靠地判断生成内容的知识归属源。作者团队推出了 AttriWiki 自监督管线,它无需人为构建知识冲突,便能自动化生成高质量的带标注训练数据。利用 AttriWiki 训练的探测器在 Llama-3.1-8B、Mistral-7B 和 Qwen-7B 等代表性模型上取得了高达 0.96 的 Macro-\(F_1\) 评分,并能无缝迁移至 SQuAD 与 WebQuestions 等基准测试(取得 0.94–0.99 的 Macro-\(F_1\) 分数),同时展现出卓越的零样本泛化能力。此外,该研究重点指出,知识归因失配会急剧抬高模型的输出错误率,这更加印证了在大语言模型中建立先进知识归因分析框架的迫切性。

This paper demonstrates that simple linear probes trained on hidden model representations can reliably identify the knowledge source of a generation. The authors introduce AttriWiki, a self-supervised pipeline that automatically generates high-quality labeled training data without relying on knowledge conflicts. Probes trained using AttriWiki achieve up to 0.96 Macro-\(F_1\) across models like Llama-3.1-8B, Mistral-7B, and Qwen-7B, transfer seamlessly to benchmarks such as SQuAD and WebQuestions (0.94–0.99 Macro-\(F_1\)), and generalize zero-shot to prior benchmarks. Furthermore, the study highlights that attribution mismatches drastically elevate error rates, emphasizing the critical need for advanced knowledge attribution frameworks in LLMs.


元数据与参考信息

Metadata & Reference Information


获取途径与资源

Access & Resources