Residual stream

The residual stream is the running vector a transformer carries from one layer to the next. Each attention block and each feed-forward block reads that vector, computes an update, and adds the update back. The stream at the top of the network is the sum of the embedding and every update along the way. When people say they found a refusal direction "in the residual stream," they mean they read that running vector at a chosen layer and found a direction in it that tracks refusal.

Why measurements happen there

The stream is a single place where the model's current state is visible. You do not have to pick one attention head and hope. A direction that matters for the next token tends to show up as a component of this vector, because the next token is read from it. Refusal, on the account in Arditi et al. 2024, is mediated by one such direction. That result is why abliteration is even a small edit instead of a full retrain: if the behavior lives on one axis of this stream, you can aim at the axis.

Reading it is not the same as editing it

Reading the stream tells you the direction. Editing the weights changes whether later tokens get pushed along it. Ablation and orthogonalization are the edits. A probe that only reads the stream, and never changes a weight, is a measurement tool. It can tell you a model is refusing. It does not produce an abliterated checkpoint. Hosted inference on Refuseless does not expose residual-stream activations. You send messages and receive a completion. The stream is inside the worker for the length of that forward pass, which is a useful picture and not an API feature.

If you are implementing the edit yourself, the layer you choose matters. A direction that is clear at one layer can be muddled at another. Published recipes name their layers. Copying a vector from one model family into another is not a method. The refusal vector page is the thing you are looking for inside the stream.