Research map / Large language models / Language, safety and interpretability
Multilingual Unlearning in LLMs: Transfer, Dynamics, and Reversibility
Chaoyi Xiang, Olga Ohrimenko, Benjamin I. P. Rubinstein, Lea Frermann
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| unlearn in a language sharing family and script with the query language (English to German) | DPO unlearning · TOFU multilingual forget set, Qwen2.5-7B | NLI score change | Improved |
| unlearn English and query a same-family language (Russian) instead of a distant one (Chinese) | DPO unlearning · TOFU multilingual forget set, Qwen2.5-7B | NLI score change: -9 ± 2 (RU) vs -4 ± 5 (CH) | Improved |
| unlearn a low-coverage language (English in Turkish-fine-tuned model) to affect the fine-tuned language | DPO unlearning · FT Turkish, query Turkish, Qwen2.5-7B | NLI score change: -55 ± 8 | Improved |
| use high-coverage language (English, Chinese) as the unlearning source rather than German, Russian or Turkish | DPO unlearning · cross-lingual transfer, Qwen2.5-7B | transfer strength | Improved |
| instruct the model to answer in the fine-tuning language (cross-lingual prompting) | Qwen2.5-7B unlearning evaluation · FT English, query German | NLI score gain: +61 | Improved |
| unlearn with DPO on small forget set | cross-lingual alignment of hidden states · Qwen2.5-7B, layer-wise cosine similarity | cosine similarity early and middle layers | No clear effect |
| unlearn with DPO | representation of forget questions · PCA of hidden states | separation from base model | Mixed |
| inject extracted steering vector at inference (subtract scaled unlearning direction over layers l..l+2) | unlearned Qwen2.5-7B · English forget set | NLI score recovery: recovers over half of lost performance | Improved |
| inject steering vector extracted only from English data | unlearned Qwen2.5-7B · non-English forget sets | NLI score recovery | Improved |
| inject isotropic Gaussian random vector instead of extracted steering vector (ablation) | unlearned Qwen2.5-7B · forget set | NLI score recovery | No clear effect |
| use Gemma2-9B instead of Qwen2.5-7B for steering recovery | steering vector injection · forget set | NLI score recovery: 50% (Qwen) -> 90% (Gemma) recovered | Improved |
| replace DPO with gradient ascent or NPO (ablation) | multilingual unlearning · cross-lingual transfer and steering | transfer pattern | No clear effect |
Closest papers by meaning
- Do LLMs Really Forget? Evaluating Unlearning with Knowledge Correlation and Confidence Awareness (NeurIPS 2025 · similarity 0.92)
- Wisdom is Knowing What not to Say: Hallucination-Free LLMs Unlearning via Attention Shifting (NeurIPS 2025 · similarity 0.92)
- In-Context Unlearning: Language Models as Few-Shot Unlearners (ICML 2024 · similarity 0.91)
- Towards Effective Evaluations and Comparisons for LLM Unlearning Methods (ICLR 2025 · similarity 0.91)
- A Closer Look at Machine Unlearning for Large Language Models (ICLR 2025 · similarity 0.91)
- Rethinking LLM Unlearning Objectives: A Gradient Perspective and Go Beyond (ICLR 2025 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
