Base versus chat: what RLHF changes inside — repurposes the instruments of → Steering doubles as interpretability

Nothing about the technique changes between these two uses, only the question it is pointed at. Inter-layer cosine similarity was first built to ask how a single behavior's representation evolves with depth inside one model, finding that vectors from nearby layers resemble each other closely and that this similarity decays more slowly in the network's second half (contrastive-activation-addition, §"8.2 Similarity between vectors generated at different layers", p. 7). The base-versus-chat comparison repurposes the identical metric, cosine similarity between vectors, and the identical transfer test, applying one layer's vector at another, to ask a different question entirely: not how a behavior's representation changes with depth, but how it changes with RLHF. The similarity between base-model and chat-model vectors peaks specifically between layers 7 and 15, and base-extracted vectors still steer the chat model's behavior most strongly in that same range (contrastive-activation-addition, §"8.3 Comparing representations between base and chat models", p. 8). The instruments were never behavior-specific; they were depth-of-representation instruments all along.