💬 NLP

Persona-Guided LLM Agents for Task-Oriented Dialogue

This paper introduces a training-free framework demonstrating that large language model agents can effectively adapt to user personalities in task-oriented dialogues to improve satisfaction and task completion, while revealing a trade-off with truthfulness that is best resolved through cue-based adaptation rather than explicit knowledge.

Original authors: Maryam Shoaeinaeini, Brent Harrison, A. B. Siddique

Published 2026-08-20

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Maryam Shoaeinaeini, Brent Harrison, A. B. Siddique

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a conversation where you are not just asking for information, but also hoping to feel understood. In the world of artificial intelligence, this is the frontier of task-oriented dialogue. For years, computer programs designed to help us book hotels or find restaurants have focused almost entirely on getting the job done: finding the right dates, checking the prices, and confirming the reservation. They are efficient, but they often feel robotic, ignoring the way a person speaks, their mood, or their personality. Meanwhile, other researchers have shown that large language models—the powerful AI systems behind many modern chatbots—can mimic human personalities in open-ended chats, acting shy, bold, or friendly when asked. But a critical question remained unanswered: can these models keep a specific personality while still solving a concrete problem without losing their way? Does adjusting to a user's personality actually make the interaction better, or does it just distract the machine from its goal?

To find the answer, researchers at the University of Kentucky built a controlled experiment where two artificial intelligence agents talked to each other. One agent played the role of a customer with a specific personality, while the other acted as the helpful system assistant. They tested this setup across thousands of simulated conversations about booking hotels and finding restaurants. The goal was to see if the system could successfully complete the task while adapting to the customer's style, and whether that adaptation made the experience more satisfying. They explored three different ways the system might know about the customer's personality. In the first scenario, the system knew nothing about the user's character. In the second, the system had to guess the user's personality just by listening to how they spoke. In the third, the system was told the user's personality explicitly, like having a secret file on the customer before the conversation even began.

The results revealed a delicate balance between being helpful and being human. The researchers found that the AI agents could indeed express distinct personalities while still completing their tasks successfully. The system agents managed to finish the booking or search goals in nearly ninety percent of the cases, proving that adding a layer of personality did not break the machine's ability to work. However, not all personalities were equally easy to express. Traits like being outgoing or agreeable came through clearly in the dialogue, while traits like being introverted or closed-off were much harder for the AI to portray consistently. The system could mimic a chatty user well, but it struggled to capture the quiet reserve of a more withdrawn one.

When it came to the quality of the interaction, adapting to the user's personality generally improved the experience. The systems that adapted, whether by guessing or by being told, were better at satisfying the specific constraints of the request and providing the right information. Users in the simulation rated these personalized interactions as more satisfying than those where the system remained neutral. Yet, this improvement came with a cost. The more the system tried to match the user's personality, the more likely it was to make small, unsupported claims or lose a bit of its strict grounding in the facts. This created a trade-off: the system became more personable and responsive, but slightly less truthful.

The study also uncovered a surprising nuance about how the system learns about the user. When the system was explicitly told the user's personality, its performance depended heavily on how clearly that personality was shown. If the user acted very strongly according to their trait, the system's adaptation was excellent. But if the user's personality was subtle or hard to pin down, the explicit instructions sometimes led the system to overreact or misunderstand. In contrast, the system that had to infer the personality from the conversation itself proved to be the most reliable. It did not need the user to act out a specific role perfectly; it simply listened to the cues in the dialogue and adjusted accordingly. This approach offered the best overall balance, improving satisfaction without the risks associated with relying on explicit labels.

Ultimately, the research suggests that the future of helpful AI lies not in rigid programming or in blindly following a personality profile, but in the ability to listen and adapt in real time. The most effective systems are those that can read the room, picking up on the subtle shifts in how a person speaks, and adjusting their tone to match without losing sight of the task at hand. This approach allows the machine to feel more like a natural conversational partner while still remaining a competent tool, bridging the gap between cold efficiency and warm interaction.

Technical Summary: Persona-Guided LLM Agents for Task-Oriented Dialogue

Problem Statement

While prior research has established that Large Language Models (LLMs) can express diverse personality traits in open-ended text generation, it remains unclear whether they can maintain these traits within Task-Oriented Dialogue (TOD) without compromising task completion. Furthermore, the impact of adapting a system's behavior to a user's personality on interaction quality is not well understood. In TOD, systems must satisfy strict schema constraints, gather slot values, and execute valid API calls to achieve user goals. There is a risk that adapting too strongly to personality cues could reduce grounding, distort recommendations, or interfere with task completion. This paper addresses the tension between personalization (adapting to user style) and task-grounding (adhering to schema constraints).

Methodology

The authors propose a training-free framework to simulate and evaluate personality-aware TOD interactions. The study utilizes the Schema-Guided Dialogue (SGD) dataset, focusing on Hotel and Restaurant domains.

Experimental Setup

Agents: The framework simulates interactions between an LLM-based User Agent and an LLM-based System Agent.
Models Evaluated: GPT-4o, Qwen3-Next-80B, and Gemini 2.0 Flash.
Personality Traits: The study covers the Big Five traits and their opposite poles (10 total traits: Extraversion, introversive, Agreeableness, disagreeable, Conscientiousness, unconscientious, Neuroticism, stable, Openness, and closed).
Conditions: To isolate the effect of personality access, three system conditions are tested:
1. Neutral: The system receives no personality information.
2. Try: The system must infer the user's personality from dialogue cues.
3. Oracle: The system is explicitly given the user's target trait.
Generation: Dialogues are generated using one-shot prompting without fine-tuning. The user agent is instructed to express the target trait implicitly through behavior, while the system agent must remain schema-compliant.

Evaluation Metrics

The generated dialogues are evaluated across four dimensions:

Task Outcomes: Dialogue Completion and Task Success (schema-constrained task completion).
Personality Realization: Measured via a TOD Trait Score (dialogue-level expression) and a Big Five Inventory (BFI) questionnaire probe (prompt-level alignment).
System Quality: Evaluated by an LLM judge (Gemma-4-31B) on Constraint Satisfaction, Inform Rate, and Truthfulness (1–5 scale).
Pairwise User Satisfaction: Evaluators compare paired dialogues to determine which system interaction is more satisfying.

Key Findings

1. Feasibility of Personality-Aware TOD (RQ1)

LLM agents can jointly express personality and maintain strong task performance. The framework achieved 89.94% Dialogue Completion and 85.27% Task Success on average. However, trait realization is uneven:

Traits like Extraversion, Agreeableness, Neuroticism, Openness, and disagreeable were expressed clearly (high TOD Trait Scores).
Traits like introversive and closed were significantly harder to realize reliably in multi-turn TOD, despite strong alignment in the separate BFI probe.

2. Impact of Personality Access on Quality and Satisfaction (RQ2)

Adapting to user personality generally improves system quality and user satisfaction but introduces a trade-off with truthfulness:

Improvements: Both Try and Oracle conditions significantly outperformed Neutral in Constraint Satisfaction and Inform Rate.
Trade-off: Both adaptive conditions significantly reduced Truthfulness compared to Neutral, confirming a personalization–grounding trade-off.
Optimal Strategy: The Try condition (cue-based adaptation) provided the best overall balance. It achieved the highest average system-quality scores and was preferred over both Neutral and Oracle in pairwise satisfaction.
Specific Trait Dynamics:
- Oracle gains were significant for socially clear traits (e.g., Extraversion, Openness) but could fail for difficult traits. For disagreeable users, Oracle settings caused models (specifically Qwen3 and Gemini) to mirror the user's hostile tone, leading to uncooperative responses and lower satisfaction.
- Try was more robust for difficult traits (e.g., disagreeable, introversive), avoiding the mirroring artifact seen in Oracle.

3. Dependence on Trait Realization Strength (RQ3)

The benefit of personality adaptation depends on how the system accesses personality information:

Oracle: Gains in satisfaction are positively correlated with the strength of trait realization ( $\rho = +0.685$ ). Explicit conditioning is most effective when the user's dialogue clearly exhibits the target trait.
Try: Gains are largely insensitive to realization strength ( $\rho = -0.103$ ). Because the system adapts to observed cues rather than a predefined label, it provides stable benefits regardless of how strongly the trait is expressed.

Significance and Claims

The paper claims to provide the first controlled, training-free evaluation of personality adaptation in schema-guided TOD. Its primary contributions are:

Demonstrating Feasibility: Showing that LLMs can express personality in goal-directed dialogue without sacrificing task completion, though some traits are harder to realize than others.
Identifying a Trade-off: Revealing that personalization improves constraint satisfaction and user satisfaction but can degrade truthfulness (grounding).
Proposing a Robust Strategy: Arguing that cue-based adaptation (Try) is a more reliable route to personality-aware TOD than explicit conditioning (Oracle). It avoids the risks of over-amplifying difficult interaction styles and remains effective even when trait realization is weak.
Methodological Contribution: Providing a framework that isolates the effect of personality access while keeping domain structure and task requirements fixed.

The authors conclude that while explicit personality labels can be beneficial when traits are clearly expressed, cue-based adaptation offers a more robust path for personality-aware TOD without the need for fine-tuning. Future work is suggested to validate these findings with human users and explore hybrid strategies.