arXiv 2502.01636
Lifelong Knowledge Editing requires Better Regularization
By Akshat Gupta, Phudish Prateepamornkul, et al.
Published 2025-02-03
Mindmap
Browse the paper's core ideas, clusters, and relationships in a structured outline.
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 interna…