dorsal/arxiv
View SchemaEvaluating Cross-Lingual Unlearning in Multilingual Language Models
| Authors | Tyler Lizzo, Larry Heck |
|---|---|
| Categories | |
| ArXiv ID | 2601.06675vv1 |
| URL | https://arxiv.org/abs/2601.06675 |
| License | http://creativecommons.org/licenses/by/4.0/ |
Abstract
We present the first comprehensive evaluation of cross-lingual unlearning in multilingual LLMs. Using translated TOFU benchmarks in seven language/script variants, we test major unlearning algorithms and show that most fail to remove facts outside the training language, even when utility remains high. However, subspace-projection consistently outperforms the other methods, achieving strong cross-lingual forgetting with minimal degradation. Analysis of learned task subspaces reveals a shared interlingua structure: removing this shared subspace harms all languages, while removing language-specific components selectively affects one. These results demonstrate that multilingual forgetting depends on geometry in weight space, motivating subspace-based approaches for future unlearning systems.
{
"annotation_id": "bed48d50-40a5-4351-a0d6-adec5dc22dd9",
"date_created": "2026-02-17T05:53:08.615000Z",
"date_modified": "2026-02-17T05:53:08.615000Z",
"file_hash": "9ae25e30a2df0849ce4a7f91fdeccb12a5871e62308201bbc49b2f199a4b5037",
"private": false,
"record": {
"abstract": "We present the first comprehensive evaluation of cross-lingual unlearning in multilingual LLMs. Using translated TOFU benchmarks in seven language/script variants, we test major unlearning algorithms and show that most fail to remove facts outside the training language, even when utility remains high. However, subspace-projection consistently outperforms the other methods, achieving strong cross-lingual forgetting with minimal degradation. Analysis of learned task subspaces reveals a shared interlingua structure: removing this shared subspace harms all languages, while removing language-specific components selectively affects one. These results demonstrate that multilingual forgetting depends on geometry in weight space, motivating subspace-based approaches for future unlearning systems.",
"arxiv_id": "2601.06675",
"authors": [
"Tyler Lizzo",
"Larry Heck"
],
"categories": [
"cs.CL"
],
"license": "http://creativecommons.org/licenses/by/4.0/",
"title": "Evaluating Cross-Lingual Unlearning in Multilingual Language Models",
"url": "https://arxiv.org/abs/2601.06675",
"version": "v1"
},
"schema_id": "dorsal/arxiv",
"source": {
"execution_id": "894127c2-b189-4f30-9a63-0b9620cbeef9",
"id": "arXiv Dataset",
"type": "Model",
"variant": "snapshot-2026-01-17",
"version": "0.1.0"
},
"user_id": 1000002
}