Contrastive Activation Addition (CAA) — skips the head search required by → Inference-Time Intervention (ITI)

CAA and ITI extract their vectors the same way — Li et al.'s ITI "employ[s] the same Mean Difference vector extraction approach as CAA" (contrastive-activation-addition, §"2 Related work", p. 2) — which makes their difference entirely about where and how broadly the vector is applied, not how it's computed. ITI first uses linear probes on a contrastive question-answering dataset to identify "a sparse set of 'truthful' attention heads," then shifts only those heads' activations along the true/false MD direction at inference. CAA skips the probe-and-search step entirely: "our technique can be applied directly to the residual stream without searching for individual attention heads" (contrastive-activation-addition, §"2 Related work", p. 2). That simplification is paid for by generality rather than against it — CAA is validated on a broader range of alignment-relevant behaviors (seven, versus ITI's single truthfulness axis) and on RLHF-trained chat models rather than a base model. So the same extraction formula scales from a narrow, head-localized, probe-dependent intervention (ITI) to a broad, stream-wide, probe-free one (CAA) once the localization step is dropped.