arXiv cs.CL· Zesheng Shi, Yucheng Zhou, Jing Li·· 12 小时前AI 评分32
约束式知识遗忘（CKU）：通过保留有用知识实现更安全的 LLM 对齐
Safety Alignment via Constrained Knowledge Unlearning
AI 导读
研究者提出约束式知识遗忘（CKU）安全对齐策略，通过对特定 MLP 层神经元打分识别与有用知识相关的神经元子集 U，并在遗忘过程中剪除 U 中神经元的梯度，从而在缓解有害内容的同时保留有价值知识。实验显示 CKU 显著提升模型安全性且不损害整体性能，在安全与效用之间取得优于现有方法的平衡。
正文 · 原文
Abstract:Despite significant progress in safety alignment, large language models (LLMs) remain susceptible to jailbreak attacks. Existing defense mechanisms have not fully deleted harmful knowledge in LLMs, which allows such attacks to bypass safeguards and produce harmful outputs. To address this challenge, we propose a novel safety alignment strategy, Constrained Knowledge Unlearning (CKU), which focuses on two primary objectives: knowledge localization and retention, and unlearning harmful knowledge. CKU works by scoring neurons in specific multilayer perceptron (MLP) layers to identify a subset U of neurons associated with useful knowledge. During the unlearning process, CKU prunes the gradients of neurons in U to preserve valuable knowledge while effectively mitigating harmful content. Experimental results demonstrate that CKU significantly enhances model safety without compromising overall performance, offering a superior balance between safety and utility compared to existing methods. Additionally, our analysis of neuron knowledge sensitivity across various MLP layers provides valuable insights into the mechanics of safety alignment and model knowledge editing.
| Subjects: | Computation and Language (cs.CL); Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2505.18588 [cs.CL] |
| (or arXiv:2505.18588v2 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2505.18588 arXiv-issued DOI via DataCite |
Submission history
From: Shi Zesheng [view email]
[v1]
Sat, 24 May 2025 08:29:50 UTC (1,887 KB)
[v2]
Fri, 16 Oct 2026 03:24:44 UTC (1,370 KB)
来源：arXiv cs.CL · arxiv.org