Can persona consistency coexist with relevant dialogue in personalized conversation?
This explores whether an AI can stay true to a fixed personality and still respond to what you actually said — or whether holding a character forces it to ignore the conversation in front of it.
This explores whether an AI can stay true to a fixed personality and still say something relevant to the conversation at hand. The corpus suggests these two goals genuinely pull against each other — and that the tension is built into how we measure and train for persona in the first place. The clearest statement of the problem: high persona-consistency scores often come from a model simply parroting its character description back at you while ignoring what you asked Do persona consistency metrics actually measure dialogue quality?. In other words, a bot can look perfectly "in character" precisely by being unresponsive. So the first surprising finding is that consistency and relevance aren't just hard to balance — optimizing one naively can actively sabotage the other.
Why does this happen? Part of the answer is that standard training rewards per-turn quality, not coherence across a whole conversation — which is why persona consistency turns out to be roughly orthogonal to raw model capability (a far more powerful model barely outperformed a weaker one on staying in character) Does model capability translate to better persona consistency?. There's also a deeper reason the corpus keeps circling: models don't really "have" a persona to begin with. They maintain a superposition of plausible characters and sample one at generation time, so regenerating the same prompt yields a different-but-locally-consistent answer each time Do large language models actually commit to a single character?. Run a persona prompt repeatedly and the variation between runs can exceed the variation between entirely different personas — the model's own uncertainty, not stable social knowledge, is doing the steering Why do LLM persona prompts produce inconsistent outputs across runs?. If there's no fixed commitment underneath, consistency was always going to fight with responsiveness.
The more interesting half of the corpus says the coexistence is achievable — but only when persona and context are optimized *together* rather than bolted on separately. One approach models the discourse relations between turns alongside persona fidelity, so the character description and the query relevance are scored jointly instead of competing Do persona consistency metrics actually measure dialogue quality?. Another borrows a trick from how humans talk: give the agent an "imaginary listener" and have it check whether each utterance would actually distinguish its persona from a generic alternative — this suppresses both bland and self-contradicting replies at inference time, no extra training needed Can imaginary listeners reduce dialogue agent contradictions?. A reinforcement-learning angle attacks drift directly, rewarding consistency across three scales (within a turn, across the whole conversation, and factual non-contradiction) to cut persona drift by over half Can training user simulators reduce persona drift in dialogue?.
There's also a quieter insight hiding here: maybe the problem is the *static persona list* itself. Predefined three-to-five-sentence character sheets tend to produce repetitive, contradictory dialogue, whereas personality drawn from genuine self-expression — journal-style writing that shows how a person actually talks — yields more consistent *and* more nuanced responses Why do static persona descriptions produce repetitive dialogue?. Push further and persona becomes something dynamic: an evolving intermediary between memory and action that gets re-optimized at test time against the user's recent interactions, so it tracks what the user actually wants instead of freezing at setup Can personas evolve in real time to match what users actually want?. This reframes the whole question — relevance and consistency stop competing when the persona is allowed to *update* in response to context rather than being a fixed wall the conversation has to route around.
The twist worth taking away: the very alignment training that gives models a reliable "Assistant" identity is part of what makes them rigid. Persona space turns out to be low-dimensional, dominated by a single axis measuring distance from the default Assistant mode How stable is the trained Assistant personality in language models?, and RLHF effectively locks in one communicative identity that can't switch register the way human pragmatics demands — users can't renegotiate it through conversation Can language models adapt communication style to different contexts?. So persona consistency and relevant dialogue *can* coexist — but the corpus's collective answer is that you get there by making the persona adaptive and jointly-scored with context, not by clamping a character down harder.
Sources 10 notes
High persona adherence scores often come from copying character descriptions while ignoring query relevance. MUDI jointly optimizes both by using discourse relations and graph-based coherence modeling alongside persona fidelity, showing that persona and context must be optimized together, not separately.
Claude 3.5 Sonnet achieved only 2.97% improvement over GPT 3.5 on persona consistency despite massive capability gaps, suggesting persona adherence is orthogonal to model scaling. Standard training objectives optimize for per-turn quality, not cross-turn coherence.
Shanahan's 20-questions test shows LLMs maintain a superposition of consistent objects or characters and sample from that distribution at generation time. Regenerating the same response yields different outputs, each consistent with prior context, proving no fixed commitment exists.
When the same persona prompt is run repeatedly, output variance across runs matches or exceeds variance across different personas. This reveals that model uncertainty, not stable social knowledge, drives persona-simulated outputs, making them unsuitable for simulating human annotation disagreement.
Endowing dialogue agents with an imaginary listener via Rational Speech Acts reduces persona contradiction at inference time without NLI labels or extra training. The agent simulates whether utterances would distinguish its persona from a distractor, suppressing generic or contradictory responses.
By inverting standard RL setups to train user simulators for consistency using three complementary metrics (prompt-to-line, line-to-line, Q&A consistency) as reward signals, persona drift decreases by over 55%. This approach captures distinct failure types: local drift within turns, global drift across conversations, and factual contradictions.
Journal entries capturing Big Five traits through genuine self-expression produce more consistent and nuanced dialogue than predefined 3-5 sentence persona descriptions. Personality emerges from how people express themselves, not from attribute inventories.
PersonaAgent uses structured personas to bridge episodic/semantic memory and personalized actions, optimizing them at test time by simulating recent interactions against textual feedback. Learned personas cluster meaningfully in latent space, suggesting genuine user-specific separation beyond standard post-training drift.
Research mapping hundreds of character archetypes reveals a low-dimensional persona space where the leading component measures distance from the default Assistant. Emotional and meta-reflective conversations cause predictable drift, but activation capping along this axis mitigates harmful shifts without degrading capabilities.
System prompts and RLHF training lock models into one communicative identity across all interactions, preventing the contextual register-switching and value trade-offs that characterize human pragmatics. Users cannot reshape model behavior through dialogue negotiation.