Refusal in Language Models Is Mediated by a Single Direction (2024), Representation Engineering: A Top-Down Approach to AI Transparency (2023), Programming Refusal with Conditional Activation Steering (2024), Refusal Direction is Universal Across Safety-Aligned Languages (2025)