Knowledge Editing for Large Language Models: Lifelong Updating, Multimodal Extension, and Open-Domain Generalization

Qizhou Chen

East China Normal University

2026年5月28日 星期四

下午 14:00-15:00

学院104会议室

Large language models (LLMs) have become knowledge-intensive systems, with knowledge statically encoded in their parameters while the real world evolves continuously.This mismatch raises a fundamental challenge: how can we efficiently and reliably revise erroneous, outdated, or undesirable model knowledge without retraining the entire model or disrupting unrelated capabilities? Knowledge editing provides a promising technical pathway for targeted model correction, with insights from LLMmechanistic interpretability offering guidance for identifying where editing interventions take effect.In this talk, I present my research on knowledge editing for LLMs.I identify three key limitations of existing methods: limited sustainability, insufficient modality applicability, and weak adaptability to open-domain settings. To address these limitations, I present three lines of research. First, I study lifelong knowledge editing, developing attention-based dynamic fusion and retrieval-based continuous prompt methods for long-term updates. Second, I extend knowledge editing to vision-language large models by analyzing cross-modal knowledge invocation and proposing attribution-guided editing frameworks.Third, I investigate open-domain knowledge editing by constructing a unified large-scale benchmark and further proposing an information-bottleneck-based editing framework to improve cross-domain adaptability and generalization robustness.These studies advance knowledge editing from single-shot, language-only, and low-complexity settings toward lifelong, multimodal, and open-domain complex scenarios, contributing to the development of more reliable, controllable, and trustworthy LLMs.