nucleation

LIT-1998: Persona Vectors: Monitoring and Controlling Character Traits in Language Models

1 min read

Properties5
StatusActive — read 2026-11-03 (NOTE-1524, Skimmed: main text in full, appendices C.2, G.2 and M only), from arXiv v3; worth reading as the method behind PSM's interpretability evidence: one difference-of-means trait direction tracks prompt-induced and finetuning-induced persona shifts, with r = 0.76–0.97 for the latter in two 7–8B models. It is a handle on movement along named traits, not a fitted persona distribution.
Version2
Tagsrepresentation-learning, self, anthology-candidate
Cited as “paper” byNOTE-1524 — Chen, Arditi, Sleight, Evans & Lindsey — Persona vectors
Cited as “source” byTHEORY-1313 — In two 7–8B chat models, the shift that finetuning induces along a difference-of-means persona vector correlates with how strongly the finetuned model then expresses that trait, r = 0.76–0.97 across synthetic datasets for evil, sycophancy and hallucination, above cross-trait baselines of 0.34–0.86; the last-prompt-token projection predicts a response's trait at r = 0.75–0.83 overall but mainly by telling prompt types apart, 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

Graph View

Chen, R., Arditi, A., Sleight, H., Evans, O., Lindsey, J. (2025). Persona Vectors: Monitoring and Controlling Character Traits in Language Models. arXiv preprint. arXiv:2507.21509.

Read on 2026-11-03, main text in full (NOTE-1524, Skimmed; THEORY-1313, THEORY-1334).

Key takeaways

  • A persona vector is a per-trait difference-of-means direction found by contrastive system prompting; the trait must be promptable.
  • Finetuning shifts along it track trait expression (r = 0.76–0.97; cross-trait 0.34–0.86); prompt-time projections mostly tell prompt types apart (within-condition r as low as 0.245).
  • Negative traits and humor shift together; penalizing the shift fails as the trait moves to other directions.
  • No distribution over personas is fitted. Its practice content (steering, data screening) is the anthology’s.

Standing in the record

Directions in activation space that monitor and steer traits; the nearest empirical handle on a persona distribution.

Backlinks