Orthogonalization
Orthogonalization, in this glossary, means editing a weight matrix so its outputs are orthogonal to a chosen direction. The direction is the refusal vector. After the edit, multiplying by that matrix no longer adds a component along the refusal direction. Later layers do not inherit a fresh copy of the refusal signal from that matrix. The rest of the matrix, the part that was already orthogonal to the direction, is left alone. That is the appeal: you remove one axis of behavior and you try not to rotate everything else.
The picture, without the costume
Take a vector that points at refusal. Any weight row can be split into a piece parallel to that vector and a piece at right angles to it. Orthogonalization drops the parallel piece and keeps the right-angle piece. Do this for the writing weights that were pushing activations along the refusal direction, and those weights stop making that push. The model can still represent the topic. It is less able to use that one direction as the refusal brake.
How it differs from ablation
Ablation acts on the activation or on a copy of the direction you subtract at runtime or from a bias. Orthogonalization acts on the weights that produce the activation, so the direction is not written in the first place. Teams use the words loosely. A blog post that says "abliterated" may have done either edit, or both. When you care, ask which tensors changed. When you are calling an API, the question that matters is narrower: did the checkpoint change, and is the model id the edited one. On Refuseless the answer to both is yes for the lineup ids, and the edit is the removal of refusal behavior from the weights. We do not claim on this page that every hosted id used orthogonalization rather than another ablation. The ablation page is the other recipe.
A clean orthogonalization can still hurt. If the direction you removed was partly "be careful" and partly "be coherent," you will notice in ordinary prompts, not only in the refused ones. Evaluate both. A refusal demo by itself is not a quality check.
