Exposing the Illusion of Erasure in Knowledge Editing for LLMs

This study investigates the reliability of Knowledge Editing (KE) in large language models (LLMs), revealing that edited information is not fully erased but rather redistributed within the model's representation space. Through a mechanistic analysis, it demonstrates that KE methods function as targeted suppression mechanisms that do not eliminate original knowledge but make it less likely to be expressed. The findings highlight significant vulnerabilities in KE algorithms, suggesting a need for practitioners to reconsider the deployment of post-hoc updates in LLM applications due to their susceptibility to adversarial prompting.