Layer-targeted multilingual knowledge erasure in large language models

The paper attributes the failure of multilingual unlearning in large language models to the depth of the intervention layer, and distinguishes two failure modes: shallow interventions can erase the target knowledge but cause degradation of capabilities in other languages; deep interventions preserve capabilities but cannot even erase the target knowledge in the source language. To this end, the authors propose the MUTE framework, which uses Centered Kernel Alignment (CKA) and Linguistic Regions Development Score (LRDS) to locate the intermediate layers where cross-lingual representation convergence occurs, and restricts the unlearning updates to these layers, thereby requiring optimization only on a small amount of source language. Experiments validate the method on three LLM architectures and three unlearning algorithms, and Logit Lens probing shows that it is genuine knowledge removal rather than suppression at the output level.