Reproduce “Identifying Knowledge Editing Types in Large Language Models” on OpenResearch

About this paper

Xiaopeng Li, Shasha Li, Shangwen Wang, Shezheng Song, Bin Ji, Huijun Liu, Jun Ma, Jie Yu

This research introduces the Knowledge Editing Type Identification (KETI) task and KETIBench dataset to identify whether and how knowledge in Large Language Models (LLMs) has been altered. Experiments show baseline identifiers can achieve an average F1 score of 0.745, distinguishing between benign and malicious edits and demonstrating generalization across editing methods.

Key insights

• Identifying distinct types of knowledge edits in LLMs is feasible, with baseline identifiers achieving an average F1 score of 0.745 in distinguishing between benign and malicious changes.
• Baseline identifiers exhibit considerable cross-domain generalization, capable of classifying edit types even when trained on different editing methods, suggesting potential robustness against unknown manipulation techniques.
• The ability to detect an edit's type is not strongly dependent on the full success of the knowledge editing method, implying that identifiers can detect the *intent* of an edit from residual traces.

Summary by alphaXiv.