PCA visualization of contrastive activations — screens datasets for → CAA steering vector construction (Mean Difference over multiple-choice contrast pairs)

Before CAA extracts a steering vector from a dataset, the paper first checks whether that dataset's activations are even separable by the behavior at all. The PCA visualization step projects a contrastive dataset's residual-stream activations at the answer-letter token and checks for behavioral clustering -- separation by whether the model's output matches the target behavior -- which the authors call useful "for determining whether a dataset will enable the generation of effective steering vectors" (contrastive-activation-addition, §"3.2 Visualizing activations for contrastive dataset analysis", p. 4). That makes visualization a screening step that precedes construction rather than an afterthought: caa-steering-vector-construction's Mean Difference procedure averages a difference in activations over an entire dataset at a chosen layer (contrastive-activation-addition, §"3 Method", p. 3), but that average is only meaningful once behavioral clustering confirms the representation being differenced actually tracks the behavior rather than just the prompt's surface format. A dataset showing only letter clustering, with no behavioral split, is a warning sign the resulting vector may encode the wrong thing entirely.