This illustrates a widespread problem affecting large language models (LLMs): even when an English-language version passes a safety test, it can still hallucinate dangerous misinformation in other ...

There’s a shocking lack of understanding of the physics underlying this commonplace phenomenon, but researchers are on the ...