1 paper across 1 session
Refusal directions in LLMs work across languages, revealing shared jailbreak mechanisms and raising the need for stronger multilingual safety.
