Multilingual Amnesia: On the Transferability of Unlearning in Multilingual LLMs

TL;DR

This paper investigates the transferability of unlearning in multilingual large language models, focusing on data and concept unlearning across ten diverse languages, revealing insights into stability and cross-lingual effects.

Contribution

It extends unlearning benchmarks to ten languages and analyzes how linguistic similarities influence unlearning transfer in multilingual LLMs.

Findings

Unlearning is more stable in high-resource languages.

Asymmetric transfer effects occur between related languages.

Syntactic similarity predicts cross-lingual unlearning behavior.

Abstract

As multilingual large language models become more widely used, ensuring their safety and fairness across diverse linguistic contexts presents unique challenges. While existing research on machine unlearning has primarily focused on monolingual settings, typically English, multilingual environments introduce additional complexities due to cross-lingual knowledge transfer and biases embedded in both pretraining and fine-tuning data. In this work, we study multilingual unlearning using the Aya-Expanse 8B model under two settings: (1) data unlearning and (2) concept unlearning. We extend benchmarks for factual knowledge and stereotypes to ten languages through translation: English, French, Arabic, Japanese, Russian, Farsi, Korean, Hindi, Hebrew, and Indonesian. These languages span five language families and a wide range of resource levels. Our experiments show that unlearning in high-resource languages is generally more stable, with asymmetric transfer effects observed between typologically related languages. Furthermore, our analysis of linguistic distances indicates that syntactic similarity is the strongest predictor of cross-lingual unlearning behavior.