文章背景与核心概要
随着大语言模型(LLM)在隐私保护和版权合规方面的需求日益增长,“机器遗忘”(Unlearning)技术应运而生,旨在从已训练好的模型中安全地抹去特定敏感信息。然而,现有的研究大多停留在简单的表面指标恢复评估上,忽视了遗忘操作背后更深层的结构性原因。本文深入探讨了知识纠缠(Knowledge Entanglement)——即某个事实与模型内部其他信息的结构连接关系——如何决定该事实在经历遗忘操作后是否会残留或泄露。
通过对多种不同领域和模型规模的遗忘算法(如 WHP 和 GA+KL)进行全面分析,研究人员发现遗忘并非仅仅切断输出,而是直接作用于底层知识结构。更重要的是,通过操纵提示词的纠缠得分,作者证实了其具有因果支配作用,并据此提出了一种新型的模型审计方法,能够精准预测哪些信息在遗忘后极易发生“泄露”。这一发现为未来构建更具鲁棒性的安全大模型提供了重要的理论支撑和实用工具。
What the "Spotless" Mind Remembers: How Knowledge Entanglement Shapes What Leaks After Unlearning in LLMs
Authors: Aakriti Shah, Yifan Hu, Thai Le
arXiv: 2510.25732 [cs.CL]
Summary
This research investigates the phenomenon of "unlearning" in Large Language Models (LLMs), moving beyond simple recovery metrics to examine the structural causes of information leakage. The authors propose that a fact's knowledge entanglement—its structural connection to other information within the model—is a primary predictor of whether that fact will persist after an unlearning attempt.
By analyzing various unlearning algorithms (WHP and GA+KL) across multiple domains and model sizes, the study reveals that unlearning does not merely remove outputs but acts upon the underlying knowledge structure. The authors demonstrate that by manipulating a prompt's entanglement score, they can causally influence recall, enabling a new method for auditing models to predict which information is likely to "leak" post-unlearning.
本研究调查了大语言模型(LLM)中的“遗忘”(unlearning)现象,超越了简单的恢复指标,进而检查信息泄露的结构性原因。作者提出,事实的知识纠缠(knowledge entanglement,即其与模型内部其他信息的结构连接)是预测该事实在遗忘尝试后是否会持续存在的主要指标。
通过分析跨多个领域和模型规模的各种遗忘算法(WHP 和 GA+KL),该研究表明遗忘不仅是删除输出,而且是对底层知识结构起作用。作者证明,通过操纵提示词的纠缠得分,他们可以因果性地影响模型召回率,从而实现一种用于审计模型的新方法,以预测哪些信息在遗忘后可能会“泄露”。
Key Findings
- Entanglement as a Predictor: Before unlearning, highly entangled facts are recalled more frequently by the model (\(r = +0.39\) to \(+0.51\)).
- Algorithmic Divergence:
- WHP (Weight-Harmonized Pruning): Weakens the correlation between entanglement and recall but maintains a positive relationship.
- GA+KL: Consistently inverts this relationship (\(r = -0.14\) to \(-0.25\)), a novel finding in unlearning literature.
- Causal Evidence: Through direct manipulation of prompt entanglement scores, the researchers confirmed that unlearning affects the internal knowledge structure rather than just the surface-level output.
- Practical Application: The study introduces a predictive model capable of auditing LLMs to identify and "triage" prompts that are at high risk of leaking information after unlearning.
核心发现
- 纠缠作为预测指标: 在遗忘之前,高度纠缠的事实被模型更频繁地召回(\(r = +0.39\) 到 \(+0.51\))。
- 算法分歧:
- WHP(权重对齐剪枝,Weight-Harmonized Pruning): 削弱了纠缠与召回之间的相关性,但仍保持正向关系。
- GA+KL: 持续颠覆这种关系(\(r = -0.14\) 到 \(-0.25\)),这是遗忘文献中的一项新发现。
- 因果证据: 通过直接操纵提示词纠缠得分,研究人员证实遗忘影响的是内部知识结构,而不仅仅是表层输出。
- 实际应用: 该研究引入了一个能够审计 LLM 的预测模型,以识别和“分诊”在遗忘后具有高信息泄露风险的提示词。
Paper Metadata
| Attribute | Details |
|---|---|
| Submitted | 29 Oct 2025 (v1), 26 Aug 2026 (v2) |
| Subjects | Computation and Language (cs.CL); Artificial Intelligence (cs.AI) |
| ACM Classes | I.2.7; I.2.6; I.2.4; G.2.2 |
| Resources | View PDF | HTML | TeX Source |
论文元数据
属性 详情 提交时间 2025年10月29日 (v1), 2026年8月26日 (v2) 学科分类 计算与语言 (cs.CL);人工智能 (cs.AI) ACM 类别 I.2.7; I.2.6; I.2.4; G.2.2 资源链接 查看 PDF | HTML | TeX 源码
License
许可证