Persona vector — specializes for character traits → Steering vector

A persona vector is a steering vector scoped to a specific kind of target — personality and character traits rather than an open list of alignment-relevant behaviors — and validated for roles the generic concept was not originally built around. The paper studies three main traits (evil, sycophancy, hallucination) plus four more in Appendix G (optimism, impoliteness, apathy, humor), and for each confirms not only that steering along the vector shifts behavior, but that the same vector reads behavior it did not cause: projecting a prompt's final-token activation onto the persona direction predicts the trait expression of a response not yet generated, correlating at r = 0.75–0.83 across system-prompt and many-shot elicitation (persona-vectors, §"3.3 Monitoring prompt-induced persona shifts via projection", p. 5). Beyond that pre-generation reading, persona vectors are also injected during training itself, not only at inference, through preventative steering (persona-vectors, §"5.2 Preventative steering limits behavioral shifts during finetuning", p. 8). So where the general steering-vector concept covers a direction added or subtracted during a single forward pass, persona vectors extend the same object into a monitor that operates before generation begins and a regularizer that operates during weight updates — two roles beyond the inference-time nudge the concept is otherwise centered on.