THEORY-1334: Persona shifts induced by prompting and by finetuning are mediated by the same linear trait directions, so finetuning on data that implies a trait moves the model along the trait's persona vector as a system prompt would; Chen et al. suspect personas are latent factors that persist across many tokens

Source

Chen et al. (2025), LIT-1998; read in NOTE-1524.

What was actually shown

• The same vector, extracted from system-prompted contrasts, steers the trait (§3.2), tracks prompt-induced shifts (§3.3) and tracks finetuning-induced shifts (§4.2, THEORY-1313).
• Subtracting it after finetuning lowers the trait (§5.1); adding it during finetuning prevents much of the shift (§5.2).
• A loss penalizing change along it is “ineffective”, “We suspect this is because the optimization pressure pushes the model to represent the personality trait using alternative directions” (p. 9).
• The authors’ explanation of the generalization: “we suspect it has to do with personas being latent factors that persist for many tokens” (pp. 13–14).

What this does not say

• That the direction is necessary for the trait: the regularization result suggests it is not.
• That training is conditioning in PSM’s strict sense: the paper does not compare finetuned models to conditioned ones beyond the shared direction.