Refusal in Language Models Is Mediated by a Single Direction

arxiv.org

saved by: 0 people
related read on: Pirate Face - The permanence layer for sovereign AI from the discussion