Activation engineering / representation engineering (technique family) — empirically motivated the formalization of → Linear representation hypothesis

The paper places this relationship in a specific order: practice came before theory. Section 2 closes its literature review by noting that, "beyond steering behaviors, work on activation engineering has also motivated a formalization of 'linear representation' (Park et al., 2023)" (contrastive-activation-addition, §"2 Related work", p. 3) — naming Park et al.'s hypothesis as a response to the family's accumulating results (Subramani et al.'s latent steering vectors, Turner et al.'s Activation Addition, Li et al.'s ITI, Zou et al.'s Representation Engineering, Liu et al.'s In-Context Vectors), not the other way around. Each of those techniques worked, to varying degrees, by adding or shifting a single direction in activation space and getting a coherent behavioral change out — an empirical regularity that needed an explanation. The linear representation hypothesis is offered as that explanation after the fact, giving the family's scattered successes a shared theoretical account rather than supplying the theory the family was built to test. CAA itself continues the same pattern one step further: a new activation-engineering technique that gets cited as further motivation for the hypothesis it also relies on.