Tag
The paper reveals that cross-lingual harmfulness representations in large language models degrade with resource tier when evaluated using hard negatives, challenging previous conclusions that safety representations survive translation.