AlphaEdit: Null-Space Constrained Knowledge Editing for Language Models

2024-10-03

arXiv.org

AI Summary

Existing localization-based editing methods suffer from catastrophic forgetting during sequential knowledge updates. To address this, we propose Null-Space Projection Knowledge Editing (ZS-Edit): prior to gradient perturbation, the parameter update direction is orthogonally projected onto the null space of the Jacobian w.r.t. preserved knowledge—thereby guaranteeing strict invariance of original model outputs under theoretical rigor. ZS-Edit is the first method to introduce null-space projection into large language model knowledge editing, combining formal theoretical grounding with implementation simplicity (requiring only a single-line code extension). Experiments on LLaMA3, GPT2-XL, and GPT-J demonstrate that ZS-Edit improves average performance over state-of-the-art editing methods by 36.4%, while significantly enhancing edit accuracy and knowledge fidelity, and effectively mitigating catastrophic forgetting in sequential editing scenarios.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Safety and Robustness
• Search and Optimization: Learning to Search

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems

Abstract

Large language models (LLMs) often exhibit hallucinations due to incorrect or outdated knowledge. Hence, model editing methods have emerged to enable targeted knowledge updates. To achieve this, a prevailing paradigm is the locating-then-editing approach, which first locates influential parameters and then edits them by introducing a perturbation. While effective, current studies have demonstrated that this perturbation inevitably disrupt the originally preserved knowledge within LLMs, especially in sequential editing scenarios. To address this, we introduce AlphaEdit, a novel solution that projects perturbation onto the null space of the preserved knowledge before applying it to the parameters. We theoretically prove that this projection ensures the output of post-edited LLMs remains unchanged when queried about the preserved knowledge, thereby mitigating the issue of disruption. Extensive experiments on various LLMs, including LLaMA3, GPT2-XL, and GPT-J, show that AlphaEdit boosts the performance of most locating-then-editing methods by an average of 36.4% with a single line of additional code for projection solely. Our code is available at: https://github.com/jianghoucheng/AlphaEdit.

Problem

• Addresses hallucinations in large language models due to outdated knowledge
• Mitigates disruption of preserved knowledge during sequential editing
• Improves editing performance with minimal additional computational cost

Innovation

• Projects perturbation onto null space
• Preserves original knowledge during edits
• Boosts performance with minimal code addition