Post-hoc (inference-time) steering mitigation — reinstantiates → Contrastive Activation Addition (CAA)

Post-hoc steering's update rule - h_l <- h_l - av_l at every decoding step - is CAA's own intervention, h <- h + av injected at every post-prompt token (contrastive-activation-addition, §"3 Method", p. 3), with the target vector swapped from a behavioral steering vector to a persona vector and the sign flipped to suppress rather than induce. Persona vectors places itself downstream of CAA explicitly, crediting CAA and similar work with requiring 'bespoke data curation to obtain contrastive pairs' - a data-pipeline critique, not a mechanism change (persona-vectors, §"7 Related work", p. 12; §"1 Introduction", p. 1). Read backward through that lineage, post-hoc steering's headline limitation - that large coefficients trade trait suppression for MMLU accuracy (persona-vectors, §"5.1 Post-hoc steering mitigates behavioral shifts", p. 7) - is not a new discovery about persona vectors specifically. CAA had already measured essentially the same cost directly on its own seven behaviors: adding a steering vector at multiplier +-1 moved MMLU accuracy by as much as 0.06 (e.g. Corrigibility to 0.64/0.59, Hallucination to 0.64/0.57) against an unsteered baseline of 0.63 (contrastive-activation-addition, §"7 Effect of CAA on general capabilities", p. 6). In hindsight, post-hoc steering is best read as CAA's exact intervention imported wholesale, capability tax included - which is precisely why persona-vectors needed to invent preventative steering as a genuinely different mechanism rather than a better-tuned version of the inherited one.