What keeps an AI-played character from contradicting itself mid-conversation — a detailed character sheet, or something subtler?

This explores what keeps an LLM-played character coherent across a conversation when it isn't pinned to a fixed profile of traits and facts. The corpus suggests consistency doesn't come free. It comes from somewhere other than the character sheet, and where it comes from decides how well it holds.

Fixed schemas turn out to be a weak anchor. Short predefined persona lists (3-5 sentences of attributes) produce repetitive, sometimes contradictory dialogue. Personality drawn from journal entries, where someone expresses Big Five traits in their own words, gives more consistent and nuanced conversation, because personality emerges from how a person talks, not from an inventory of attributes Why do static persona descriptions produce repetitive dialogue?. High "persona consistency" scores can also be misleading. They often come from copying the character description while ignoring what the user just said, so consistency has to be optimized together with staying on topic Do persona consistency metrics actually measure dialogue quality?.

Underneath, the model isn't holding a schema at all. One view is that an LLM keeps a superposition of many characters that fit everything said so far, and each reply samples from that crowd. The conversation is the constraint, and the crowd narrows as it proceeds. That explains why regenerating a reply can change the personality while still fitting the earlier context Does an LLM commit to a single character or maintain many?. A second view covers the model's own assistant persona. Post-training installs stable dispositions that persist under adversarial pressure, unlike prompt-induced role-play that collapses under jailbreaks. Those dispositions are realized in the model rather than performed Are RLHF personas performed characters or realized dispositions? Are LLM personas realized or merely simulated through training?.

For a prompted character, though, that consistency is fragile. Run the same persona prompt repeatedly and the variance across runs matches or exceeds the variance between different personas, so model uncertainty swamps the persona Why do LLM persona prompts produce inconsistent outputs across runs?. Persona prompts mostly steer the surface of the output and leave underlying tendencies like bias unchanged Can persona prompts actually reduce bias in language models?. A much stronger model doesn't fix it. Claude 3.5 Sonnet improved persona consistency by only 2.97% over GPT 3.5, probably because standard training rewards each turn on its own rather than coherence across turns Does model capability translate to better persona consistency?.

So the workable approaches give consistency its own pressure. One trains for it directly. Multi-turn reinforcement learning on user simulators, rewarded on prompt-to-line, line-to-line and Q&A consistency, cut persona drift by over 55%. Those three signals catch drift within a turn, drift across a conversation, and factual contradictions Can training user simulators reduce persona drift in dialogue?. Another adds a different kind of structure. A three-layer persona that keeps the model in reactive response within a script produced more human-like dialogue consistency, at some cost to emotional expression Can layered persona architecture sustain coherent character behavior?. A third lets the persona keep changing. PersonaAgent treats it as a living link between memory and action, and rewrites it at test time by simulating recent interactions against feedback Can personas evolve in real time to match what users actually want?. Across all of these, consistency behaves less like a property of the character sheet and more like a property of the conversation.