How many lines of someone's real conversation does an AI need to see before it starts talking like them?

This explores how much conversation history you have to show a pretrained (base) model before it behaves like one specific person instead of a generic speaker. The corpus's most direct answer is that it takes less than you'd expect. Base models conditioned on short dialog samples predict human responses more accurately and more diversely than instruction-tuned assistants given a written persona Do pretrained models simulate humans better than instruction-tuned assistants?. No note here measures how accuracy grows from two turns to twenty. 'Short' is the finding, and the corpus has no number to attach to it.

The note credits task mismatch, not context length, for the gap. Assistant tuning optimizes a model to be a helpful assistant, and that systematically degrades its ability to imitate a person. A base model just continues text, so a few real turns work as evidence about who is speaking. This fits the view that a language model holds a superposition of possible characters. Regenerating a reply gives a different answer each time, and each answer is consistent with the context so far, so nothing is ever committed to Do large language models actually commit to a single character?. Read that way, dialog context doesn't install a persona. It narrows the set of characters the model might be sampling from. That is my inference from the two notes, not something either states.

The alternative, describing the persona in a prompt, binds much more loosely. Persona prompts change how the output reads but leave underlying bias gaps between groups unchanged Can persona prompts actually reduce bias in language models?. In annotation simulation, running the same persona prompt repeatedly produced as much variation between runs as between different personas, so the model's own uncertainty drowned out the persona signal Why do LLM persona prompts produce inconsistent outputs across runs?. A label like 'a 34-year-old nurse' gives the model little to hold onto, while a few lines of how the person actually talks give it something to condition on.

Binding a persona once doesn't keep it bound as the conversation goes on. Trained assistants sit on a dominant 'Assistant' axis in persona space, and emotional or meta-reflective conversations pull them along it in predictable ways How stable is the trained Assistant personality in language models?. User simulators drift too, and training them with multi-turn RL against consistency rewards cut that drift by over 55% Can training user simulators reduce persona drift in dialogue?. A lighter fix at inference time is to have the agent imagine a listener and check whether an utterance would distinguish its persona from a distractor. That reduced contradictions without extra training Can imaginary listeners reduce dialogue agent contradictions?. Other work keeps the persona current instead, rewriting it at test time by simulating recent interactions against feedback Can personas evolve in real time to match what users actually want?.

The kind of context matters as much as the amount. Realistic synthetic dialogue needed subtopic, Big Five persona traits and 11 contextual characteristics working together, and that combination reached 90.48% of in-domain performance Can synthetic dialogues become realistic through layered diversity?. Conditioning a simulator on a session-level profile plus turn-level intent also produced conversations that were hard to tell from real ones Can controlled latent variables make LLM user simulators realistic?. The picture that emerges is a short, well-chosen sample of real dialog to bind the persona, plus something to keep it from drifting as the conversation runs.