Edited Knowledge Fusion with Large Language Models
Yaqin Wen · 2025
Recent research in knowledge editing has demonstrated significant potential for updating large language models with new memories, enabling the replacement of outdated information and the integration of specialized knowledge. In this work, we introduce a novel task, Knowledge Editing Fusion, which focuses on merging multiple independent edits within the same model without prior aggregation or re-editing. To address this challenge, we propose the Multi-Edit Weight Fusion (MEFusion) method for parameter-modifying editing. MEFusion leverages weight deltas to integrate edits efficiently.