Year: 2026
Type: conference-paper
Abstract: Ensuring that Large Language Models (LLMs) remain useful throughout their lifetimes requires frequent updates to maintain factual accuracy, yet standard "locate-then-edit" methods suffer from catastrophic interference in sequential settings, where successive updates degrade previously integrated knowledge.In this work, we propose EvoEdit, a framework that achieves stable, large-scale sequential editing through sequential null-space alignment.Unlike previous works, EvoEdit dynamically projects knowledge edits into the null space of both original and previously modified knowledge, with theoretical guarantees of output invariance across extended edit sequences.This approach balances model adaptability with knowledge preservation, preventing the cumulative interference that commonly leads to model collapse with previous methods.Empirical evaluations across multiple LLMs demonstrates state-of-the-art performance at moderate scales (e.g., 2K edits) and provides substantial improvements in stability and retention at large scales of up to 10K edits.Furthermore, the method achieves up to a 3.5× speedup over existing baselines, establishing EvoEdit as a principled, efficient, and scalable solution for reliable knowledge management in real-world LLM deployments.Code is available at here.
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7166813849