Generalizable Lifelong Model Editing via Preference Optimization

2026-09-29

AI Summary

This study addresses the limitations of existing lifelong model editing methods, which are prone to overfitting and consequently suffer from restricted knowledge generalization and degraded general capabilities. To mitigate these issues, this work proposes GLIME, a framework that introduces Direct Preference Optimization (DPO) into the lifelong editing setting for the first time. By integrating an experience replay mechanism with gradient projection constraints, GLIME effectively balances the synergistic updating of both new and previously acquired knowledge. Experimental results demonstrate that GLIME significantly enhances knowledge generalization in lifelong scenarios on large language models while maintaining high editing accuracy and preserving overall model utility. Ultimately, this framework establishes a novel paradigm for sustainable knowledge updating in foundation models.

Abstract

Knowledge editing enables rapid updates of specific factual knowledge in large language models (LLMs) without full retraining. However, more realistic scenarios call for a lifelong framework that handles continual updates rather than one-off modifications. In such settings, existing editing methods often overfit to target prompts, significantly degrading both the generalization of the edited knowledge and the model's general capabilities. To address this issue, we propose GLIME (Generalizable Lifelong Model Editing), which combines knowledge editing with preference optimization over generation behavior. GLIME further incorporates replay-based editing and a gradient constraint to preserve previously edited knowledge. Experimental results show that GLIME significantly improves knowledge generalization in lifelong editing settings while maintaining both editing performance and general capabilities.

Problem

• Knowledge Editing
• Lifelong Learning
• Large Language Models
• Generalization
• Overfitting

Innovation

• Lifelong Model Editing
• Preference Optimization
• Knowledge Generalization
• Replay-based Editing
• Gradient Constraint