From Refusal Tokens to Refusal Control: Discovering and Steering Category-Specific Refusal Directions

Researchers developed a method to control AI safety refusal behavior using categorical refusal tokens in Llama 3 8B, enabling fine-grained control over when models refuse harmful versus benign requests. The technique uses steering vectors that can be applied during inference without additional training, improving both safety and reducing over-refusal of harmless prompts.

• →Categorical refusal tokens create separable directions in AI model activation space that can be extracted and controlled.
• →The steering vector approach reduces false refusals on benign prompts while maintaining safety on harmful requests.
• →The intervention method transfers across same-architecture model variants without requiring additional training.
• →A low-rank combination technique enables single controllable intervention under activation-space anisotropy.
• →The research demonstrates practical inference-time control over fine-grained AI safety behavior.