Adaptive method improves continual unlearning for safe language models
Adaptive Localization of Knowledge Negation for Continual LLM Unlearning
Machine Learning
Summary
Removing unwanted or harmful information from large language models is important to keep them safe, but doing this repeatedly can make the models less useful. The authors developed a new technique called ALKN that carefully targets only the parts of the model related to the knowledge to be removed. This keeps the rest of the model’s knowledge intact and helps maintain its usefulness, even after multiple rounds of unlearning. Tests showed that ALKN works better than earlier methods at both removing bad knowledge and keeping the model helpful.
What this means in practice
- •For ai safety teams: Safely remove harmful knowledge from deployed language models repeatedly while preserving their overall usefulness.
- •For software developers: Improve updates to AI language features by selectively editing model knowledge without degrading performance across tasks.
Authors
Abudukelimu Wuerkaixi, Qizhou Wang, Sen Cui, Wutong Xu, Bo Han, Gang Niu, Masashi Sugiyama, Changshui Zhang
Abstract
With the growing deployment of large language models (LLMs) across diverse domains, concerns regarding their safety have grown substantially. LLM unlearning has emerged as a pivotal approach to removing harmful or unlawful content while maintaining utility. Despite increasing interest, the challenges of continual unlearning, which is common in real-world scenarios, remain underexplored. Successive unlearning tasks often lead to intensified utility degradation. To effectively unlearn targeted knowledge while preserving LLM utility, it is essential to minimize changes in model parameters by selectively updating those linked to the target knowledge, thereby ensuring other knowledge remains unaffected. Building on the task vector framework, we propose a new method named ALKN (Adaptive Localization of Knowledge Negation), which uses dynamic masking to sparsify training gradients and adaptively adjusts unlearning intensity based on inter-task relationships. Comprehensive experiments across three well-established LLM unlearning datasets demonstrate that our approach consistently outperforms baseline methods in both unlearning effectiveness and utility retention under continual unlearning settings.