What is refusal ablation actually doing inside a language model? Our new preprint shows that standard refusal ablation can be interpreted as a minimum-confidence latent-space evasion: it projects the representations onto the boundary of a refusal probe. Can we push beyond it?

Language models encode refusal as a manifold expressed through multiple, closely related directions. In our latest #AAAI26 paper, with @RaffaeleMura3, we bridge LLM safety and interpretability by answering the question: 👉 Do LLMs encode refusal behavior as a manifold?