Refusal in Language Models Is Mediated by a Single Direction

Research identifies a single mechanistic direction in LLM activations that controls refusal behavior, offering a tractable target for alignment interventions.

Categories: Research

Read at source: https://arxiv.org/abs/2406.11717