nucleation

LIT-1991: The Assistant Axis: Situating and Stabilizing the Default Persona of Language Models

1 min read

Properties5
StatusActive — read 2026-11-03 (NOTE-1505, Skimmed: main text in full, appendices not read), from arXiv v1; worth reading for the geometry of persona space in three open-weight models, the Assistant at one end of a pre-existing first axis, and for the finding that within a conversation the persona's position follows the latest user message, which bears on any episode-level model of personas (CLAIM-536, QUESTION-370).
Version2
Tagsrepresentation-learning, self, anthology-candidate
Cited as “paper” byNOTE-1505 — Lu, Gallagher, Michala, Fish & Lindsey — The Assistant Axis
Cited as “source” byTHEORY-1330 — In Gemma 2 27B, Qwen 3 32B and Llama 3.3 70B, activation vectors for 275 prompted character roles span a low-dimensional space whose first principal component runs from fantastical to Assistant-like roles, with role loadings correlated above 0.92 between models and the default Assistant at its extreme; the contrast vector from the default Assistant to the role mean aligns with it and is largely present before post-training, THEORY-1339 — In multi-turn conversations a chat model's persona is not fixed for the episode: its position along the Assistant Axis is set mainly by the most recent user message (ridge R² 0.53–0.77 from message embeddings) rather than by where it was before (R² 0.10 for the change), so that therapy-like and AI-philosophy conversations drift away from the Assistant and coding and writing ones do not

Graph View

Lu, C., Gallagher, J., Michala, J., Fish, K., Lindsey, J. (2026). The Assistant Axis: Situating and Stabilizing the Default Persona of Language Models. arXiv preprint; PMLR v306. arXiv:2601.10387.

Read on 2026-11-03, main text in full (NOTE-1505, Skimmed; THEORY-1330, THEORY-1339).

Key takeaways

  • Persona space is low-dimensional (4–19 PCs for 70%); PC1 loadings agree across models (r > 0.92), with the Assistant at one end.
  • The axis is largely inherited from pre-training; post-training adds its AI-assistant associations.
  • Within a conversation the position is predicted by the latest user message (R² 0.53–0.77), not by the previous position (R² 0.10).
  • Activation capping is practice, the anthology’s.

Standing in the record

A direction along which the default Assistant persona is situated and stabilized; bears on what post-training installs (TERM-072’s note of 2026-10-22, NOTE-1069).

Backlinks