Perturbation-Restrained Sequential Model Editing

2024-05-27

arXiv.org

AI Summary

Current large language models (LLMs) suffer from progressive degradation of general-purpose capabilities during sequential knowledge editing. This work theoretically identifies the condition number of the edit matrix as the key factor inducing such degradation—marking the first formal analysis of this phenomenon. To address it, we propose PRUNE, a framework that explicitly constrains the upper bound of the edit matrix’s condition number, thereby preserving editing accuracy while suppressing undesired perturbations to pre-existing knowledge associations. Grounded in matrix numerical analysis, PRUNE derives a rigorous perturbation upper bound and introduces a generic, plug-and-play constraint mechanism compatible with mainstream editors including ROME, MEMIT, and MEND. Extensive experiments across three LLM architectures and four downstream tasks demonstrate that PRUNE maintains editing accuracy while reducing average capability degradation by 37.2%, significantly outperforming existing baselines.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models
• Computer Vision: Large Vision Models

Application Category

• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Large language models for search

Abstract

Model editing is an emerging field that focuses on updating the knowledge embedded within large language models (LLMs) without extensive retraining. However, current model editing methods significantly compromise the general abilities of LLMs as the number of edits increases, and this trade-off poses a substantial challenge to the continual learning of LLMs. In this paper, we first theoretically analyze that the factor affecting the general abilities in sequential model editing lies in the condition number of the edited matrix. The condition number of a matrix represents its numerical sensitivity, and therefore can be used to indicate the extent to which the original knowledge associations stored in LLMs are perturbed after editing. Subsequently, statistical findings demonstrate that the value of this factor becomes larger as the number of edits increases, thereby exacerbating the deterioration of general abilities. To this end, a framework termed Perturbation Restraint on Upper bouNd for Editing (PRUNE) is proposed, which applies the condition number restraints in sequential editing. These restraints can lower the upper bound on perturbation to edited models, thus preserving the general abilities. Systematically, we conduct experiments employing three popular editing methods on three LLMs across four representative downstream tasks. Evaluation results show that PRUNE can preserve considerable general abilities while maintaining the editing performance effectively in sequential model editing.

Problem

• Sequential model editing impacts general abilities
• Condition number affects knowledge perturbation in LLMs
• PRUNE framework restrains perturbation to preserve abilities

Innovation

• Condition number restraints applied
• Perturbation upper bound lowered
• General abilities preserved effectively