Unveiling the Pitfalls of Knowledge Editing for Large Language Models

• DOI: 10.48550/arxiv.2310.02129
• Publication Date: 2023-01-01
• Authors: Zhoubo Li; Ningyu Zhang; Yunzhi Yao; Mengru Wang; Xi Chen; Huajun Chen
• Keywords: Unintended consequences; Code (set theory)

As the cost associated with fine-tuning Large Language Models (LLMs) continues to rise, recent research efforts have pivoted towards developing methodologies edit implicit knowledge embedded within LLMs. Yet, there's still a dark cloud lingering overhead -- will editing trigger butterfly effect? since it is unclear whether might introduce side effects that pose potential risks or not. This paper pioneers investigation into pitfalls for To achieve this, we new benchmark datasets and propose innovative evaluation metrics. Our results underline two pivotal concerns: (1) Knowledge Conflict: Editing groups of facts logically clash can magnify inherent inconsistencies in LLMs-a facet neglected by previous methods. (2) Distortion: Altering parameters aim factual irrevocably warp innate structure Experimental vividly demonstrate inadvertently cast shadow unintended consequences on LLMs, which warrant attention future works. Code data are available at https://github.com/zjunlp/PitfallsKnowledgeEditing.