Storage Is Not Strategy: State-Conditioned Support Control for LLM Unlearning
作者: Tianhao Qian, Ziming Hong, Chongyang Gao, Kezhen Chen, Lixu Wang
分类: cs.LG, cs.CL
发布日期: 2026-09-29
备注: 18 pages
💡 一句话要点
提出状态条件支持控制以优化LLM的去学习策略
🎯 匹配领域: 支柱九:具身大模型 (Embodied Foundation Models)
关键词: 大型语言模型 去学习 干预评分 动态重排序 优化策略 数据隐私 机器学习
📋 核心要点
- 现有的LLM去学习方法在优化过程中固定参数子集,导致干预效果不佳,无法适应动态变化。
- 本文提出干预评分和选择性动态干预重排序,通过动态评估干预效果来优化去学习过程。
- 实验结果显示,所提方法在多个基准测试中均表现优于现有方法,提升了去学习的有效性。
📝 摘要(中文)
许多现有的大型语言模型(LLM)去学习方法在优化过程中选择固定的小参数子集。然而,与目标最相关的参数并不一定是最佳更新参数,候选干预的效果也会随着优化过程而变化。本文引入干预评分(Intervention Score),通过预测实际去学习更新的效果来对可编辑组进行排名,并形成静态干预值基线(Static-IV)。随后,提出选择性动态干预重排序(DIR-R),仅在校准探测器证明比较合理时重新审视该子集。实验结果表明,该方法在多个基准测试中表现优异,支持了将定位、初始干预选择和检查点依赖的支持修订分开的观点。
🔬 方法详解
问题定义:本文旨在解决现有LLM去学习方法中固定参数子集导致的干预效果不佳的问题。现有方法未能适应参数更新过程中的动态变化,影响了去学习的效果。
核心思路:论文提出干预评分(Intervention Score)来动态评估可编辑组的效果,并引入选择性动态干预重排序(DIR-R),以便在合适时机重新审视参数子集,从而优化去学习过程。
技术框架:整体架构包括干预评分计算、静态干预值基线(Static-IV)建立和选择性动态干预重排序(DIR-R)三个主要模块。首先计算干预评分,然后形成基线,最后在校准探测器的指导下进行动态重排序。
关键创新:最重要的创新在于引入了干预评分和动态重排序机制,使得干预选择不再是静态的,而是能够根据优化过程的变化进行调整。这一方法与传统的固定参数子集方法有本质区别。
关键设计:在干预评分的计算中,考虑了潜在的附带损害,并通过优化算法动态调整干预参数。此外,采用了多种基准测试来验证方法的有效性,确保了实验结果的可靠性。
🖼️ 关键图片
📊 实验亮点
实验结果显示,所提方法在Natural-TOFU数据集上在19/20的比较中表现出积极的描述性边际,并在LACUNA定位精度基准测试中,所有六个负偏好优化(NPO)和SimNPO比较中均表现出更高的终端效用,NPO边际范围为+0.431至+0.848,SimNPO边际范围为+0.503至+0.571。
🎯 应用场景
该研究的潜在应用领域包括大型语言模型的训练与优化、机器学习模型的去学习过程以及数据隐私保护等。通过优化去学习策略,可以在保护用户隐私的同时,提升模型的性能和适应性,具有重要的实际价值和未来影响。
📄 摘要(原文)
Many localized large language model (LLM) unlearning methods select a small parameter subset from a localization signal and keep it fixed during optimization. The parameters most associated with a target, however, need not be the best ones to update, and candidate interventions can change value as optimization proceeds. In a controlled experiment, a storage-localization score reaches an area under the receiver operating characteristic curve (AUROC) of 0.981, yet storage identity agrees with the better intervention on only 17/36 targets, while low-rank adaptation (LoRA) wins 35/36. We introduce Intervention Score, which ranks editable groups by the predicted effect of the actual unlearning update while accounting for collateral damage, and use it to form the static intervention-value baseline (Static-IV). We then introduce selective dynamic intervention re-ranking (DIR-R), which revisits that subset only when a calibrated probe justifies the comparison. On the Natural-TOFU dataset, our method has positive descriptive margins in 19/20 comparisons between methods and objectives, although several are near zero. On the LACUNA localization-precision benchmark, our mean terminal utility is higher in all six negative preference optimization (NPO) and SimNPO comparisons: NPO margins range from +0.431 to +0.848, and SimNPO margins range from +0.503 to +0.571. The gradient-difference (GradDiff) objective reveals substantial field dependence. Relative to Static-IV, the primary four-field GradDiff evaluation has six wins, six ties, and no losses, with mean and median paired gains of +0.165 and +0.0025. The evidence supports separating localization, initial intervention selection, and checkpoint-dependent support revision.