Tag
This paper studies multilingual unlearning in LLMs by extending the TOFU benchmark to five languages. It finds that unlearning transfer varies by script and family, operates primarily in later decoding layers, and that a single steering direction can recover much of the suppressed knowledge across languages.