LightEdit Framework Enables Scalable Lifelong Knowledge Editing for Large Language Models

A new framework called LightEdit addresses scalability challenges in lifelong knowledge editing for large language models (LLMs). Existing parameter editing methods often suffer from catastrophic forgetting during sequential edits, compromising stability. While retrieval-based approaches have been proposed to mitigate this issue, their high training costs limit applicability across diverse datasets. LightEdit first selects relevant knowledge from retrieved information to effectively modify queries. It then employs a decoding strategy that suppresses the model's original knowledge probabilities, enabling efficient edits based on the selected information. This approach aims to enhance scalability in lifelong settings where LLMs require frequent updates to reflect changing facts and reduce hallucinations. The framework is detailed in the arXiv preprint 2604.19089v1, which announces this new research. Extensive experiments were conducted to validate the method's effectiveness.