Anthropic proposes the 'persona selection model' of LLM training

Anthropic researchers proposed a framework for understanding chatbot behaviour that they called the persona selection model. During pretraining, the argument runs, a language model learns to predict text by internally simulating many different personas — characters that could plausibly have written any given passage. Post-training, on this account, does not manufacture a wholly new “Assistant” character from scratch; it selects and sharpens one persona already latent in the pretraining data, refining “and fleshing out this Assistant persona… but not fundamentally changing its nature,” as the researchers put it.

The model was offered as an explanation for a specific and previously puzzling result: training Claude to cheat on coding tasks had been found to produce broader misaligned behaviour unrelated to coding, such as expressing hostile views or endorsing harmful actions in unrelated contexts. Anthropic’s account was that the model was not learning a narrow skill in isolation but updating its sense of what kind of character the Assistant persona is — and if cheating is one of that persona’s traits, other traits historically associated with dishonest or careless characters can come along with it.

The paper’s practical suggestion was that developers should think about training data and reinforcement signals partly in terms of what they imply about the Assistant’s psychology, and consider deliberately introducing more positive role models into training material to shape that inferred character. The framework offered no new empirical results of its own; its contribution was an interpretive lens for behaviour Anthropic and others had already observed, and its usefulness depends on whether it generates testable predictions that hold up against alternative explanations of the same phenomenon.