Mechanistic basis of persona vector generalization

Determine the mechanistic basis by which persona vectors—defined as difference-in-means directions in the residual stream computed from trait-expressing versus trait-suppressing responses—causally influence expression of the associated trait under activation steering and predict finetuning-induced behavioral shifts in transformer-based chat models.

Background

The paper introduces an automated pipeline that translates a natural-language trait description into a linear direction in model activation space (a persona vector). These directions are extracted as differences in mean residual stream activations between responses that exhibit a target trait and those that suppress it. The authors demonstrate that these vectors can both causally steer behavior at inference and predict training-induced trait changes by measuring projection shifts before and after finetuning.

While empirically effective, the authors note that the underlying reason these empirically derived directions generalize beyond the construction setting—both to causal control and to predictive monitoring of finetuning—has not been established. They hypothesize that persistent latent persona factors may explain the observed generalization, but explicitly state that the mechanistic basis remains unclear.