Steering outside the residual stream (proposed extension) — would localize representations beyond → Residual stream

Steering outside the residual stream is a proposal to trade the residual stream's convenience for its ambiguity. The residual stream is a running sum: every attention block and MLP writes its output into the same accumulating vector, so an intervention added into the stream at layer L perturbs the combined total of everything written into it so far, without distinguishing which sub-component's contribution is being altered. The suggested extension is to intervene "at other points in the model, such as after the MLP but before merging into the residual stream" (contrastive-activation-addition, §"9.1 Suggested future work", p. 9) — a site where the perturbation acts on one specific computation's output rather than the accumulated total. The stated payoff is diagnostic as much as practical: by moving the intervention to different pre-merge points, "we could learn more about where representations are localized in the model and achieve more targeted effects" (contrastive-activation-addition, §"9.1 Suggested future work", p. 9). The residual stream's summing structure is exactly what makes CAA's current intervention easy to apply and hard to localize.