Evaluating Cross-Lingual Unlearning in Multilingual Language Models

Tyler Lizzo, Larry Heck · 2026

Read Evaluating Cross-Lingual Unlearning in Multilingual Language Models by Tyler Lizzo, Larry Heck (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-06675.txt.

Description

Tyler Lizzo and Larry Heck evaluate cross-lingual unlearning in multilingual models in 2026. Unlearning aims to remove specific facts from a trained model. Theirs is the first comprehensive evaluation across languages, they report.

They use translated TOFU benchmarks in seven language and script variants. Major unlearning algorithms are tested on those translated fact sets. Most fail to remove facts outside the training language, they find. Utility often remains high even while forgetting fails across languages. Subspace projection consistently outperforms the other methods they tested. It achieves strong cross-lingual forgetting with minimal degradation, they report. Privacy and compliance teams should note the central warning. Deleting a fact in English may leave it intact elsewhere.

Catalog record

CREATORTyler Lizzo, Larry Heck
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE51 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2601-06675

Browse the full library · About the Tower of Babel