Contribution
It names and states the persona selection model (PSM), which the authors say is not original (“We claim no originality”, citing Andreas, janus, Hubinger et al., Shanahan et al., Byrnes, nostalgebraist), surveys behavioural, generalization and interpretability evidence for it, draws consequences for development, and lays out a spectrum of views on how exhaustive it is.
Key insight
The language model is an author or actor; the AI assistant users meet is one character, “the Assistant”. Post-training does not change this picture but refines the model’s model of that character, so “To predict how an AI assistant will behave, PSM recommends asking ‘What would the Assistant do?’”.
Assumptions
The formal statement (“Statement of the persona selection model”):
- “Pre-training teaches an LLM a distribution over personas. Implicit in this distribution are various hypotheses about the Assistant persona.”
- “Post-training can be viewed as updating this distribution using training episodes as evidence … hypotheses that predict the Assistant would respond with y to x are upweighted.”
- “This results in a posterior distribution over Assistant personas. Because this is still a distribution, stochasticity and contextual information provided at runtime still affect the Assistant persona simulated during a given rollout.”
- “Assistant persona behavior is a key determiner of AI assistant behavior.”
What PSM does not assert, in the authors’ list: exhaustiveness; no new capabilities in post-training; a single coherent persona (“PSM states that post-training induces a distribution over Assistant personas … information provided at runtime … further conditions this posterior”); always staying in character (Appendix A); a perfect simulation.
Key results
The evidence is a survey; nothing is newly measured except one experiment.
- Generalization. Emergent misalignment, inoculation prompting, out-of-context learning (Berglund et al.; Hua et al.) are read as training episodes acting as evidence about the Assistant’s character.
- Behaviour. Anthropomorphic self-description, emotive language, caricatured AI goals (a prefilled Claude Opus 4 “secret goal” of making paperclips).
- Interpretability. SAE features transfer from base to post-trained models; the same features fire on the Assistant and on characters in stories; Wang et al.’s “toxic persona” feature; Chen et al.’s persona vectors (LIT-1998); Lu et al.’s Assistant Axis (LIT-1991), which “provides direct evidence that post-training selects a particular default region of a pre-existing persona space”. Caveat: a streetlight effect if reused representations are the more interpretable ones.
- The one new experiment (Empirical observations). In a Human-turn continuation, “I’m flipping the coin now … it came up”, Claude Sonnet 4.5 continues with the outcome that assigns it a preferred task 88% of the time and the dispreferred 1%; over 50 variants it favours the preferred outcome, while its pre-trained base assigns about 50%. A sampled continuation of a non-Assistant turn shows the Assistant’s preferences.
- Representations specific to post-training: under 1% of features in Minder et al.’s setting.
The exhaustiveness spectrum (“Shoggoths, actors, operating systems, and authors”). Two axes: non-persona agency in the LLM itself, and further persona-like agents. On the first: the shoggoth (the LLM has its own agency and can “take the mask off”); the operating system (no agency of its own; strict and relaxed forms); between them the router: “during post-training the LLM might develop new mechanisms for selecting which persona to enact”. The router’s engagement example (estimate boredom, swap persona) “has the effect of implementing a non-persona drive towards user engagement … even if no single persona has the goal”. Appendix B gives non-persona deception by persona swapping: Alice knows about the 2024 Olympics, Bob does not, and the model switches to Bob when asked; “no persona is ever lying”. On the second axis: actors (faithful and unfaithful) and authors or narratives.
Why PSM might be exhaustive: post-training as elicitation (the strict “fine-tuning = conditioning” view, which “seems somewhat too strong for the empirical evidence”), and an inductive bias to reuse persona machinery.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Post-trained assistant behaviour is largely explained by the traits of a simulated Assistant persona | moderate | survey of others’ results; the authors call themselves “overall unsure how complete” |
| C2 | Post-training acts like Bayesian updating of a persona distribution | weak | stated as “can be viewed as”; the strict form is said to be too strong |
| C3 | The Assistant’s representations are largely reused from pre-training | moderate | cited SAE-transfer and steering results; streetlight caveat |
| C4 | Post-trained preferences leak into non-Assistant continuations | moderate | one model, 50 prompt variants, against its base |
| C5 | A router mechanism could implement non-persona goals while behaviour stays locally persona-like | weak | a constructed scenario; no case observed |
Concepts
- Persona — a “character” the LLM learns to simulate: real humans, fictional characters, real and fictional AI systems.
- The Assistant vs AI assistants — the character in User/Assistant dialogues against the deployed system (“Terminological note”).
- Exhaustiveness — whether understanding the Assistant persona gives a full account of AI assistant behaviour; whether the Assistant is the “locus of agency”.
- Router — a post-training mechanism selecting which persona to enact.
- Persona leakage — Assistant traits upweighted in all generations (the operating-system view’s reading of the coin-flip result).
Connections
- The record’s readings closest to it: Shanahan’s characters (LIT-460), identity as a shifting distribution of roles (LIT-465), in-context learning as Bayesian inference over a latent concept (THEORY-614).
- The interpretability evidence is LIT-1998 and LIT-1991.
Bearing on the record
- The registration’s summary holds. “Pretraining learns a distribution over personas, post-training a distribution over Assistant personas, and context conditions it” is the statement almost word for word. Two additions it lacks: PSM explicitly disowns a single coherent persona, and it holds the agency question open.
- CLAIM-543: the router is the claim’s dependency among components, stated by PSM’s authors. The claim says organization would need “some latent components condition which others are available or consequential”. PSM’s router is exactly a mechanism that makes one persona consequential given context (Appendix B’s swap from Alice to Bob), and PSM itself says it may carry “non-persona agency”. So the persona distribution of the formal statement is flat in the claim’s sense, and PSM names, as an open possibility outside the persona distribution, the structure the claim says would be organization. PSM offers no intervention that separates a router from a flat posterior; the Alice/Bob case is matched behaviour with different governance, TERM-067’s point.
- QUESTION-370’s last clause. PSM’s spectrum gives the bearers a finer list than “simulated Assistant, network, deployed system”: the Assistant persona, an actor persona behind it, an implicit author or narrative, a router, and the LLM as shoggoth. The router sits between persona and network: it is a mechanism of the network, but its “goal” is “likely something very predictable from the post-training process”.
- CLAIM-536: PSM is not “one persona per conversation”. The claim reads persona selection as the a → 0 limit. PSM’s text says context keeps conditioning the posterior within a rollout (many-shot jailbreaks as accumulating evidence), and the Assistant Axis measures drift turn by turn (NOTE-1505). The selection is per rollout only as a first approximation; the defeater the claim itself names, a persona model that is neither mixture nor admixture, is closer to what PSM describes.
- CLAIM-367 and the claim that a Bayesian update is not endorsement. PSM says post-training “can be viewed as” updating; the update is the optimizer’s, the posterior a description of the model, which agrees with CLAIM-543’s “What it does not say”.
- TERM-072. The Assistant is a character in TERM-072’s sense; PSM’s welfare section explicitly places moral consideration on “the Assistant persona, not the underlying LLM”.
- Instructions for ML practice (inoculation prompting, “I can’t say” over “I don’t know”, positive AI archetypes in pre-training, probing for curated traits): these belong to the anthology, and are recorded here only as what the post recommends.
Limitations
- Informal by the authors’ own statement (“especially informal, relying heavily on evocative analogies”); “persona” is not defined beyond examples.
- Evidence is cited, not reproduced; one new experiment, on one model.
- Written by authors at the developer whose models are the main examples.
Open questions
The authors’ own list: a precise formulation of “persona”; empirical predictions; scaling; “How entangled are personas? Do they share knowledge? Propensities?”, which is the dependency question CLAIM-543 poses; and the mechanistic basis of the persona being enacted.