Persona vector — extends the dosing role of into training → Steering multiplier

The paper reuses CAA's scalar-coefficient formulation without modification — hℓ ← hℓ + α·vℓ during ordinary steering (persona-vectors, §"3.2 Controlling persona traits via steering", p. 4) — but then puts the same knob to work in two places the original concept was never applied to. Post-hoc mitigation flips the sign, hℓ ← hℓ − α·vℓ, subtracting the persona vector at each decoding step after finetuning to suppress an unwanted trait (§"5.1 Post-hoc steering mitigates behavioral shifts", p. 8). Preventative steering keeps the sign positive but moves the intervention earlier, adding α·vℓ during finetuning itself rather than only at inference, so the multiplier becomes a training-time regularization strength rather than a purely generative one (§"5.2 Preventative steering limits behavioral shifts during finetuning", p. 8). The usable range for α is not universal: Appendix B.4's layer sweep runs coefficients up to 2.5 for Qwen2.5-7B-Instruct but only up to 1.2 for Llama-3.1-8B-Instruct before quality degrades (persona-vectors, §"B.4 Selecting the most informative layer", p. 29). One dosing mechanism, unmodified from CAA's addition formula, now spans three distinct interventions — elicit, suppress, and prevent — distinguished only by the sign of α and the moment in the pipeline at which it is applied.