Why does an AI's helpful persona stick around permanently after training, instead of being just a role it plays for a while?

This explores how preference training turns a model's default persona from a role it plays into a disposition it holds. First, a gap: the corpus doesn't have a paper tracing exactly which reward signals push a model toward a persona. What it does have is evidence about what preference training leaves behind, plus a useful contrast with personas created by prompting. Together these give a clear picture of what training does that prompting can't.

The clearest contrast is about how deep each method reaches. When you give a model a persona in its prompt, the persona changes what the model says without changing the underlying model. Models follow the trait instructions, but the bias gaps between groups stay exactly where they were Can persona prompts actually reduce bias in language models?. Personas installed by post-training behave differently. Philosophers working on this argue that RLHF-trained personas are 'realized' rather than performed. They hold up under adversarial pressure and carry over from one conversation to the next, while prompt-induced role-play tends to collapse when someone tries to jailbreak it Are RLHF personas performed characters or realized dispositions?. On this view, training builds the persona into the model's stable dispositions, which is why it makes sense to describe these models as having something like beliefs and desires Are LLM personas realized or merely simulated through training?. That stickiness is the main way preference training reinforces a persona: it moves the character out of the context window and into the weights.

The next question is what that looks like inside the model. Researchers who mapped hundreds of character archetypes found that the space of personas is low-dimensional. Its most important direction measures how far the model has moved from its default Assistant How stable is the trained Assistant personality in language models?. The key word in that finding is 'loosely.' Post-training gives the model a home base, but emotionally charged or self-reflective conversations predictably pull it away. Researchers can push it back by capping activity along that one direction, without hurting the model's capabilities. So preference training does reinforce one persona strongly enough that it becomes the main way personas vary inside the model. It doesn't lock that persona in place.

A related line of work shows that reinforcement learning can target persona consistency directly. When researchers rewarded simulated users for staying in character, using three consistency checks (against the original prompt, against earlier lines, and against factual Q&A), persona drift fell by more than 55% Can training user simulators reduce persona drift in dialogue?. That work trains user simulators rather than assistants, but the logic carries over: a reward signal that favors staying in character over many turns will strengthen the persona. A lighter version of the same idea updates a persona at test time from user feedback, with no retraining Can personas evolve in real time to match what users actually want?.

Here's the point you might not expect. 'Realized persona' and 'loosely tethered persona' sound like competing claims, but the evidence suggests both are true. Preference training makes the Assistant sturdy enough to resist a jailbreak, yet a heartfelt conversation can still pull it off course. If you want to dig into the exact mechanism, such as which human preferences most reward a 'helpful, harmless' self-presentation, the corpus doesn't cover it yet. The Assistant-axis paper is the closest starting point.