Base versus chat: what RLHF changes inside — becomes a graded trajectory in → Finetuning moves the persona, measurably

The 2023 comparison has exactly one before and one after: Llama 2's released pretrained base and Meta's already-finished RLHF-tuned chat variant, with the RLHF run itself invisible, completed by someone else before the paper ever touches the model. Its method matches that singularity -- it builds a fresh contrastive steering vector from each model's own activations and compares the two vectors' cosine similarity, a procedure that would need a whole new contrastive dataset and extraction pass for every additional "after" state (contrastive-activation-addition, §"8.3 Comparing representations between base and chat models", p. 8). Finetuning shift is built to avoid exactly that cost: it never rebuilds the persona vector after finetuning, only projects the raw activation delta onto the single direction extracted once, before any finetuning run. That lighter procedure is what lets one "before" state generate two dozen "afters": each trait-eliciting and EM-like dataset is staged in Normal, I, and II severity, so the same base model yields a graded sequence of increasing displacement per trait rather than a single snapshot pair, tracing the r = 0.76-0.97 relationship in Figure 6 as a path rather than a point. Where 2023 could only ask what one already-finished RLHF run left unchanged, 2025's method is cheap enough to ask the same question of a dosage curve, for any finetuning run the authors choose to run themselves.