Mean Difference (MD) vector extraction — supplies the extraction rule for → CAA steering vector construction (Mean Difference over multiple-choice contrast pairs)

Mean Difference is not something CAA invented; Related Work notes it was already used by Zou et al. (2023) for representation extraction and by Li et al. (2023) for Inference-Time Intervention (contrastive-activation-addition, §"2 Related work", p. 2). What CAA's steering vector construction contributes is not new arithmetic but a new interface to that shared arithmetic: it feeds MD a specific dataset (hundreds of multiple-choice contrast pairs), a specific layer L, and a specific token position (the answer letter), then plugs those into Equation 1, vMD = (1/|D|) times the sum over the dataset D of aL(p, cp) minus aL(p, cn) (contrastive-activation-addition, §"3 Method", p. 3). The paper is explicit about this: "this method of extracting the difference vector is called Mean Difference (MD)" (contrastive-activation-addition, §"3 Method", p. 3). So the relationship runs one way — MD supplies the difference-of-means machinery wholesale, and everything that makes caa-steering-vector-construction a distinct, citable procedure is decoration around that borrowed core: the dataset, the layer, the token position. A reader who assumes CAA's contribution is algorithmic misreads the paper; the contribution is entirely in what gets fed to an existing algorithm.