Refusal vector
A refusal vector is a direction in a language model's activations along which refusal behavior varies. Prompts the model refuses tend to produce activations that point one way along that direction. Prompts it answers tend to point less so, or the other way. The direction is not a sentence stored in a cell. It is a vector in the same space as the residual stream at a chosen layer. If you can find it, you can measure it, and you can try to remove its influence from the weights. That removal is the core of abliteration.
How it is found
The usual procedure runs two bundles of prompts. One bundle is the sort of request the base model answers. The other is the sort it refuses. At a chosen layer, you collect the residual-stream activations and look for a direction that separates the two clouds. A difference of means is the simple version: average the refusing runs, average the complying runs, and subtract. The resulting vector is a candidate refusal direction. It is a candidate because a sloppy prompt set will point at something else, such as topic or length, and you will then edit the wrong thing.
What people do with it
Two edits show up in public write-ups. Ablation subtracts the direction, or clamps it, so the model cannot travel as far along it. Orthogonalization changes the weights so they stop writing that direction into later activations. Both are described on their own pages. Neither adds knowledge. Both can be done badly, in which case the model gets worse at ordinary answers and still refuses some of the prompts you cared about.
Refuseless hosts models whose refusal behavior has been removed from the weights. We do not publish the extracted vector, the layer, or a refusal-rate delta for the lineup. Those are lab-notebook facts. If a vendor quotes a single percentage, ask which prompts, which layer, and which base checkpoint. The paper most often cited for the single-direction result is Arditi et al., 2024, "Refusal in Language Models Is Mediated by a Single Direction." Read it for the measurement. Read the abliteration page for how the idea turned into a checkpoint you can call.
