Related Knowledge Perturbation Matters: Rethinking Multiple Pieces of Knowledge Editing in Same-Subject

TL;DR

This paper investigates the challenges of editing multiple related pieces of knowledge for the same subject in large language models, revealing limitations of current methods and introducing a new benchmark for evaluation.

Contribution

It introduces the S2RKE benchmark for same-subject knowledge editing and analyzes the failure modes of existing locate-then-edit methods.

Findings

Locate-then-edit methods interfere when editing related knowledge pieces.

Current methods over-rely on subject information, neglecting other factors.

Proposed analysis highlights the need for more comprehensive editing strategies.

Abstract

Knowledge editing has become a promising approach for efficiently and precisely updating knowledge embedded in large language models (LLMs). In this work, we focus on Same-Subject Editing, which involves modifying multiple attributes of a single entity to ensure comprehensive and consistent updates to entity-centric knowledge. Through preliminary observation, we identify a significant challenge: Current state-of-the-art editing methods struggle when tasked with editing multiple related knowledge pieces for the same subject. To address the lack of relevant editing data for identical subjects in traditional benchmarks, we introduce the $\text{S}^2\text{RKE}$(Same-Subject Related Knowledge Editing) benchmark. Our extensive experiments reveal that only mainstream locate-then-edit methods, such as ROME and MEMIT, exhibit "related knowledge perturbation," where subsequent edits interfere with earlier ones. Further analysis reveals that these methods over-rely on subject information, neglecting other critical factors, resulting in reduced editing effectiveness.