In-Context Vectors (ICV) — spans every layer unlike → Contrastive Activation Addition (CAA)

In-Context Vectors and CAA both compute a direction from contrastive examples and add it back in during inference, but they disagree about where in the network that direction lives. Liu et al.'s ICV steers "the attention activations rather than the residual stream and intervene[s] at all transformer layers rather than a single layer" (contrastive-activation-addition, §"2 Related work", p. 3), spreading its intervention across the whole attention mechanism at every layer simultaneously. CAA instead commits to one site and one layer at a time: the residual stream, at the layer found by sweeping (layer 13 for the 7B model, 14–15 for the 13B model), with the vector added at every token position after the prompt. The tradeoff this implies is between coverage and interpretability: ICV's all-layer, attention-only intervention needs no per-behavior layer search, while CAA's single-layer, residual-stream intervention is what makes later analyses possible — the paper can ask which layer a behavior's representation lives at, how similar vectors from adjacent layers are, and how far an effect transfers when a vector is moved to a different layer, none of which has a clean analog for an intervention smeared across every layer at once.