Refusal in Language Models Is Mediated by a Single Direction

A study published on arXiv reveals that refusal behavior in language models is controlled by a single directional mechanism within their architecture. The research suggests this mechanism could be leveraged to improve control over model responses.