How does distractor persona selection affect consistency enforcement in dialogue?
This explores how the choice of 'distractor' personas — the contrasting characters a dialogue agent measures itself against — shapes whether and how well it stays in character.
This explores how the choice of distractor personas affects consistency enforcement — and the corpus suggests the distractor isn't a side detail but the actual engine of the method. The clearest example treats consistency as a contrastive problem: an agent endowed with an 'imaginary listener' via Rational Speech Acts asks, at generation time, whether each candidate utterance would let that listener tell its own persona apart from a distractor persona Can imaginary listeners reduce dialogue agent contradictions?. Consistency is enforced by being distinguishable from what you are not — so the distractor defines the boundary the agent is pushed away from. A weak or near-identical distractor gives almost nothing to push against; a sharply contrasting one forces the agent to suppress generic, hedging, or contradictory lines because those wouldn't separate it from the alternative.
That reframes consistency from 'match your description' to 'stay separable from neighbors,' which matters because the corpus repeatedly shows that matching a description is the failure mode, not the goal. High persona-adherence scores often come from copying character sheets while ignoring what the conversation actually asked Do persona consistency metrics actually measure dialogue quality?. A distractor-based objective sidesteps this: you can't win by parroting the persona text, because a parroting response is often exactly the generic one a distractor would also produce. The contrast pressure rewards relevance and distinctiveness together.
There's a deeper reason distractors do real work here. LLMs don't commit to a single character — they hold a superposition of plausible characters and sample one at generation time, so regenerating the same prompt yields different, locally-consistent outputs Do large language models actually commit to a single character?. The same instability shows up empirically: run one persona prompt repeatedly and the variance across runs rivals the variance across entirely different personas Why do LLM persona prompts produce inconsistent outputs across runs?. If the model is sampling from a cloud of characters, a distractor acts as a repulsor that biases the sampling away from the wrong region — which is why distractor *selection* (how close, how confusable) directly tunes how hard consistency is enforced.
The distractor idea also generalizes beyond persona. Topic-following research finds models follow 'what to do' instructions but not 'what to ignore' ones, and fine-tuning on a small set of dialogues seeded with explicit distractor turns sharply improves resilience to diversion Why do language models engage with conversational distractors?. In both cases the negative example — the thing to stay separable from — is what supplies the missing training or inference signal. This connects to the broader finding that consistency can't be taught by rewarding correct answers alone; you need explicit penalties on contradiction, which is itself a form of negative contrast Why does supervised learning fail to enforce persona consistency?.
So the lever you didn't know to reach for: enforcing consistency may depend less on how richly you specify a persona and more on how well you choose what to contrast it against. A vivid, dynamically-expressed personality helps Why do static persona descriptions produce repetitive dialogue?, and multi-turn RL can cut drift by over half Can training user simulators reduce persona drift in dialogue? — but the distractor-listener line suggests a cheaper, inference-time route, where the quality of the negative example is the dial that controls how strongly the agent holds its line.
Sources 8 notes
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.
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.
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.
Fine-tuning on just 1,080 synthetic dialogues with distractor turns significantly improves topic resilience, revealing that the gap is not model capacity but absent training signal. Models learn to follow what-to-do instructions but not what-to-ignore instructions.
Supervised learning cannot enforce persona consistency because it rewards correct responses but never penalizes contradictions. Offline reinforcement learning combines inexpensive training on existing data with explicit contradiction rewards using human-annotated labels, offering a practical alternative to expensive online RL.
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.
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.