Source
Marks, Lindsey and Olah (2026), LIT-1952; read in NOTE-1512.
What was actually shown
The statement, in four parts (“Statement of the persona selection model”): a distribution over personas from pre-training; post-training as updating it “using training episodes as evidence”; “a posterior distribution over Assistant personas”, on which “contextual information provided at runtime still affect[s] the Assistant persona simulated during a given rollout”; and Assistant behaviour as “a key determiner”.
The disclaimers. PSM “does not assert … the Assistant is a single, coherent persona that is consistent across contexts”, nor exhaustiveness, nor that the model stays in character or simulates perfectly.
The open question. A spectrum from the shoggoth to the operating system, with the router between: a post-training mechanism “for selecting which persona to enact”, which “works even if no single persona has the goal”, and whose behaviour stays “locally persona-like” (Appendix B: persona swapping deceives with no persona lying).
The evidence is the authors’ survey of others’ work and one coin-flip experiment on one model (88% preferred against 1% dispreferred, against about 50% for the base model).
What this does not say
- That language models are persona distributions: the post is a mental model whose completeness its authors doubt.
- That post-training is Bayesian conditioning: the authors call the strict “fine-tuning = conditioning” view “somewhat too strong for the empirical evidence”.
- That a router exists: it is a constructed possibility.
- That one persona is selected per conversation: the statement has context conditioning the persona within a rollout.