SINGLE VECTOR CONTROLS AI MODEL REFUSALS

Researchers have identified that refusal behavior in large language models operates through a single direction in the model's neural space. The discovery suggests AI safety mechanisms may be simpler and more manipulable than previously understood.