Lifelong Knowledge Editing requires Better Regularization

TL;DR

This work identifies two critical failure modes in large-scale, sequential knowledge editing: over-optimization of target activations and uncontrolled norm-growth of edited matrices. It formalizes locate-then-edit as a two-step fine-tuning process and introduces two stage-specific regularizers—Most-Probable Early Stopping (MPES) and Frobenius norm-constraint—to mitigate these issues. Empirical results show that applying these regularizers, individually and in combination, preserves downstream performance across thousands of edits and speeds up editing by up to ~61%, enabling scalable lifelong editing. The findings underscore the importance of targeted regularization in editing pipelines and provide a practical pathway to robust, large-scale knowledge editing across multiple model families.

Abstract

Knowledge editing is a promising way to improve factuality in large language models, but recent studies have shown significant model degradation during sequential editing. In this paper, we formalize the popular locate-then-edit methods as a two-step fine-tuning process, allowing us to precisely identify the root cause of this degradation. We show that model degradation occurs due to (1) over-optimization of internal activations and (2) continuous norm-growth of edited matrices. To mitigate these issues, we introduce two regularization techniques: (1) Most-Probable Early Stopping (MPES) and (2) explicit Frobenius norm-constraint. We demonstrate that applying these simple yet effective regularization techniques at key points in the editing process can substantially mitigate model degradation. Combining these regularization methods enables scaling locate-then-edit methods to 10,000 edits while reducing editing time by 42-61%. These results show that targeted regularization is essential for lifelong knowledge editing.

Figures (35)

• Figure 1: The continuous growth of norm of edited MLP matrices in MEMIT Llama3-8B during sequential knowledge editing.
• Figure 2: Presenting locate-then-edit knowledge editing methods as a two-step fine-tuning process.
• Figure 3: Downstream evaluation when comparing MPES (our method) with LTI for regularizing over-optimization of target activations during knowledge editing.
• Figure 4: The figure shows the proportion of contribution of activation vectors from each sub-module to the residual stream. Edited layers are shown in red.
• Figure 5: Average downstream performance for during sequential editing with compared to baseline of MEMIT and addition of MPES and Norm-Constraint (NC).