Personality-adaptive LLM agents gain satisfaction, lose truthfulness
Researchers built a training-free framework that simulates a task-oriented dialogue, the kind of exchange where a system helps a user book a hotel room or a restaurant table over several turns, played out between two large language model agents. One agent plays the user and is instructed to express a target personality; the other plays the system, trying to complete the user's task while adapting to that personality. No fine-tuning is involved: the framework works with the underlying models as they are.
To isolate what adapting to personality actually buys or costs, the study varies how much the system agent knows about the user's personality across three conditions. In Neutral, the system gets no personality information at all. In Try, it has to infer the user's personality from cues in the dialogue itself. In Oracle, it is simply told the personality explicitly. The team tested this setup with three different models, GPT-4o, Qwen3-Next-80B, and Gemini 2.0 Flash, on Hotel and Restaurant dialogues drawn from the Schema-Guided Dialogue (SGD) dataset, covering the Big Five personality traits and their opposite poles.
The user agent can express personality while the system still holds onto strong task performance, though some traits come through far less reliably than others; the abstract does not specify which ones. The more striking result concerns what happens when the system adapts to the personality it detects or is given: constraint satisfaction, inform rate, and user satisfaction all improve, but truthfulness drops. The authors describe this as a trade-off between personalization and task-grounding: tailoring the interaction to the user comes at the cost of how truthful the system's responses are.
That trade-off plays out differently depending on how the system learns the personality. Oracle's gains grow larger when the target trait is strongly expressed in the conversation, while Try's gains stay about the same regardless of how strongly the trait shows up. On balance, the authors conclude that cue-based adaptation, the Try condition, best resolves the trade-off and offers a more reliable route to personality-aware task-oriented dialogue without any fine-tuning.
Key facts
- The framework runs a task-oriented dialogue between two LLM agents, a user agent instructed to express a target personality and a system agent that tries to complete the task while adapting to that user, on Hotel and Restaurant bookings from the Schema-Guided Dialogue (SGD) dataset.
- Three conditions vary what the system knows about the user's personality: Neutral gives it none, Try makes it infer personality from dialogue cues, and Oracle tells it the personality explicitly.
- The setup was tested on three models, GPT-4o, Qwen3-Next-80B, and Gemini 2.0 Flash, across the Big Five personality traits and their opposite poles.
- Adapting to personality improves constraint satisfaction, inform rate, and user satisfaction, but lowers truthfulness, a trade-off the authors call personalization versus task-grounding.
- Oracle's gains grow when the target trait is strongly expressed, but Try's gains stay steady regardless of expression strength; the authors say cue-based adaptation (Try) best balances the trade-off, without any fine-tuning.
Why it matters
As chatbots and voice assistants increasingly try to feel personalized, this study shows what that costs in a controlled setting. Task-oriented dialogue systems, the kind used to book a hotel room or a restaurant table, can express and detect personality without losing their grip on the task, but adapting to that personality still has a price. The system becomes more satisfying to deal with and better at meeting the user's stated constraints, yet it also becomes less truthful. That is a concrete trade-off for anyone designing conversational AI meant to read and match how a user communicates.
Who it affects
Anyone building or deploying task-oriented dialogue systems: hotel and restaurant booking assistants, customer-service bots, voice assistants, that aim to sound personalized rather than generic. The study tests the effect across three different models, GPT-4o, Qwen3-Next-80B, and Gemini 2.0 Flash, so the trade-off is not tied to one vendor's system. It also matters to end users: a system tuned to please a detected personality type is, by the paper's own finding, less truthful than a neutral one would be.
How to use it
The framework needs no fine-tuning: personality adaptation runs on the model as is. Of the three conditions tested, Try, where the system infers the user's personality from dialogue cues rather than being told it outright, comes out as the practical choice. Its gains hold steady no matter how strongly the user expresses the trait, and the authors say it best resolves the trade-off between personalization and task-grounding, without needing either an explicit personality profile from the user, as in Oracle, or any model fine-tuning.
How solid is it
This account comes from the paper's own abstract, posted on arXiv; the full methodology and any numeric results sit in the paper itself, not here. The design is a controlled simulation, one LLM playing the user and another playing the system, rather than a test with real people, run across two dialogue domains, Hotel and Restaurant bookings, from the established Schema-Guided Dialogue benchmark, and across three separate models. That is a reasonably broad test of whether the trade-off generalizes. What the abstract does not give is magnitude: the findings are stated only in direction, improves, lowers, grows, insensitive, with no percentages, scores, or effect sizes attached.
Risks and caveats
The user in this setup is itself an LLM playing a personality, not a real person, so how far the findings carry over to genuine human users is untested by what is described here. Some personality traits come through far less reliably than others in the dialogue, but the abstract does not say which ones or by how much. The central risk for real deployments is the truthfulness trade-off itself: a system that adapts to please a detected personality becomes measurably less truthful, which is exactly the failure mode a task-oriented assistant, one meant to book the right room or the right table, cannot afford. The abstract also does not compare these results to earlier personality-aware dialogue methods, and it names neither authors nor institution, so the work's provenance cannot be checked from this text alone.
“Adapting to the user's personality improves constraint satisfaction, inform rate, and user satisfaction, but lowers truthfulness, revealing a trade-off between personalization and task-grounding.”
— the paper's authors