Persona-Guided LLM Agents for Task-Oriented Dialogue

# Persona-Guided LLM Agents for Task-Oriented Dialogue
Source: [https://arxiv.org/html/2608.18085](https://arxiv.org/html/2608.18085)
Maryam Shoaeinaeini, Brent Harrison, A\.B\. Siddique University of Kentucky Lexington, KY, USA \{maryam\.shoaei, brent\.harrison, ab\.siddique\}@uky\.edu

###### Abstract

Prior work has shown that large language models \(LLMs\) can express diverse personality traits in open\-ended text generation\. However, it remains unclear whether they can do so in a goal\-directed dialogue without compromising task completion, and whether adapting to the user’s personality improves the interaction quality\. We study these questions in task\-oriented dialogue \(TOD\), where a system helps a user accomplish a goal via multi\-turn interaction\. We build a training\-free framework that simulates a TOD interaction between two LLMs: a user agent that exhibits a target personality and a system agent that adapts to the user while completing the task\. To isolate the effect of adaptation, we vary how much the system knows about the user’s personality across three conditions\. InNeutral, the system receives no personality information\. InTry, it infers the personality from dialogue cues\. InOracle, it is given the personality explicitly\. We evaluate GPT\-4o, Qwen3\-Next\-80B, and Gemini 2\.0 Flash on Hotel and Restaurant dialogues from the Schema\-Guided Dialogue \(SGD\) dataset, across the Big Five traits and their opposite poles\. We find that the user agent can express personality while the system maintains strong task performance, although some traits are realized far less reliably than others\. 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\.Oracle’s gains grow when the target trait is strongly expressed, whereasTry’s gains are largely insensitive to realization strength\. Overall, cue\-based adaptation inTrybest resolves this trade\-off and offers a more reliable route to personality\-aware TOD without fine\-tuning\.111Code and generated dialogues will be released upon acceptance\.

Persona\-Guided LLM Agents for Task\-Oriented Dialogue

Maryam Shoaeinaeini, Brent Harrison, A\.B\. SiddiqueUniversity of KentuckyLexington, KY, USA\{maryam\.shoaei, brent\.harrison, ab\.siddique\}@uky\.edu

## 1Introduction

Large language models are increasingly used as assistants in settings where the quality of interaction matters as much as the correctness of the final answer\. In such settings, users do not only want systems that complete tasks; they also want systems that communicate in ways that feel appropriate, responsive, and personally aligned\. This has made personalization a central goal for dialogue agents, particularly in customer support, booking, and assistant scenarios where interaction style can shape user trust and satisfactionParket al\.\([2023](https://arxiv.org/html/2608.18085#bib.bib6)\); Wanget al\.\([2024](https://arxiv.org/html/2608.18085#bib.bib5)\)\.

Prior work has shown that LLMs can express diverse personality traits in open\-ended text generation\. LLMs can imitate recognizable personality patterns in questionnaires, narratives, role\-playing, and controlled generation tasksLiet al\.\([2022](https://arxiv.org/html/2608.18085#bib.bib8)\); Pan and Zeng \([2023](https://arxiv.org/html/2608.18085#bib.bib9)\); Serapio\-Garcíaet al\.\([2023](https://arxiv.org/html/2608.18085#bib.bib10)\); Jianget al\.\([2022](https://arxiv.org/html/2608.18085#bib.bib11),[2024](https://arxiv.org/html/2608.18085#bib.bib19)\); Maoet al\.\([2023](https://arxiv.org/html/2608.18085#bib.bib12)\)\. These studies suggest that LLMs can exhibit measurable Big Five traitsGoldberg \([2013](https://arxiv.org/html/2608.18085#bib.bib16)\)and maintain persona\-consistent patterns in generated text\. However, most existing evaluations study personality expression outside goal\-directed dialogue\. It therefore remains unclear whether LLMs can express personality while pursuing a concrete task, and whether adapting to the user’s personality actually improves interaction quality\.

This question is especially important in task\-oriented dialogue \(TOD\), where a system helps a user accomplish a goal through multi\-turn interaction\. Unlike open\-ended conversation, TOD systems must satisfy schema constraints, gather required slot values, issue valid API calls, and complete the user’s taskRastogiet al\.\([2020](https://arxiv.org/html/2608.18085#bib.bib13)\); Zanget al\.\([2020](https://arxiv.org/html/2608.18085#bib.bib2)\); Mosharrofet al\.\([2025](https://arxiv.org/html/2608.18085#bib.bib7)\)\. Personality\-aware behavior may improve the interaction by making the system more responsive to the user’s communication style, uncertainty, or preferences\. At the same time, it may also introduce risks: adapting too strongly to personality cues could reduce grounding, distort recommendations, or interfere with task completion\. Thus, personality\-aware TOD requires balancing personalization with task\-grounded behavior\.

Motivated by this gap, we study three research questions:RQ1: Can LLMs, without fine\-tuning, generate coherent task\-oriented dialogues that simultaneously express personality and satisfy schema constraints?RQ2: How does personality access affect system quality and pairwise user satisfaction?RQ3: How does the strength of user trait realization shape the satisfaction gains from personality conditioning, and does this relationship differ between explicit \(Oracle\) and implicit \(Try\) conditioning? To answer these questions, we build a training\-free framework for testing how personality adaptation affects LLM agents in schema\-guided task\-oriented dialogue\. We compare three system conditions:Neutral, where the system receives no personality information;Try, where it must infer the user’s personality from dialogue cues; andOracle, where it is explicitly given the user’s target trait\. This design isolates the effect of personality access while keeping domain structure, schemas, and task requirements fixed\. We evaluate GPT\-4o, Qwen3\-Next\-80B, and Gemini 2\.0 Flash by generating Hotel and Restaurant dialogues from the Schema\-Guided Dialogue \(SGD\) datasetRastogiet al\.\([2020](https://arxiv.org/html/2608.18085#bib.bib13)\)for the Big Five traits and their opposite poles\. We assess the generated dialogues along four dimensions: task outcomes, personality realization, system quality, and pairwise user satisfaction\. Figure[1](https://arxiv.org/html/2608.18085#S1.F1)summarizes the overall framework\.

![Refer to caption](https://arxiv.org/html/2608.18085v1/latex/figures/Overview1.png)Figure 1:Overview of our framework for personality\-aware task\-oriented dialogue\. A personality\-conditioned user agent interacts with a system agent under three personality\-access conditions—Neutral,Try, andOracle—within an SGD\-based API\-grounded task setting\. Generated dialogues are evaluated for task outcomes, system quality, and personality\-related behavior\.Our findings answer the three research questions as follows\. ForRQ1, LLM\-based user agents can express personality in schema\-guided dialogues while the system maintains strong task performance, although some traits are realized much less reliably than others\. RegardingRQ2, adapting to the user’s personality improves constraint satisfaction, inform rate, and pairwise user satisfaction, but reduces truthfulness, revealing a personalization–grounding trade\-off\. ForRQ3,Oraclegains increase when the target trait is strongly expressed, whereasTrygains are largely insensitive to realization strength\. Overall, cue\-based adaptation in theTrysetting offers a more reliable route to personality\-aware TOD without fine\-tuning\.

## 2Related Work

### 2\.1LLM\-based Task\-Oriented Dialogue

Task\-oriented dialogue \(TOD\) research has traditionally focused on systems that track dialogue state, fill slots, predict system actions, and ground responses in structured APIs or knowledge sources\. Progress in this area has depended heavily on high\-quality annotated datasets such as MultiWOZ and Schema\-Guided Dialogue \(SGD\), which support benchmarking across domains but also highlight the cost and brittleness of manual annotationEricet al\.\([2020](https://arxiv.org/html/2608.18085#bib.bib3)\); Zanget al\.\([2020](https://arxiv.org/html/2608.18085#bib.bib2)\); Hanet al\.\([2021](https://arxiv.org/html/2608.18085#bib.bib4)\); Rastogiet al\.\([2020](https://arxiv.org/html/2608.18085#bib.bib13)\)\. Benchmarks such as schema adherence, state tracking, and completion of user goals, which formalize the core requirements of TOD systems, remain central even when the generator is an LLM\.

Recent work has increasingly replaced fully supervised pipelines with LLM\-based approaches that use in\-context learning or reduced annotation to generalize to new domains more efficientlyMosharrofet al\.\([2025](https://arxiv.org/html/2608.18085#bib.bib7)\)\. This shift improves flexibility, but the primary focus remains task success, slot accuracy, and robustness under domain shift rather than user\-specific adaptation\. As a result, prior TOD work provides strong foundations for structured interaction, but leaves open how personality cues alter task completion, grounding, and user satisfaction in schema\-constrained settings\.

### 2\.2Personality and Persona Control in LLMs

A separate line of work studies whether large language models can express stable personality traits or personas\. Prior studies evaluate personality signals through questionnaire\-style probes, narrative generation, and controlled prompting, showing that LLMs can imitate recognizable trait patterns and exhibit measurable variation across personality dimensionsLiet al\.\([2022](https://arxiv.org/html/2608.18085#bib.bib8)\); Pan and Zeng \([2023](https://arxiv.org/html/2608.18085#bib.bib9)\); Serapio\-Garcíaet al\.\([2023](https://arxiv.org/html/2608.18085#bib.bib10)\); Jianget al\.\([2022](https://arxiv.org/html/2608.18085#bib.bib11),[2024](https://arxiv.org/html/2608.18085#bib.bib19)\)\. Related work on role\-playing and agent simulation further suggests that LLMs can maintain social roles and behavioral framing over longer interactionsWanget al\.\([2024](https://arxiv.org/html/2608.18085#bib.bib5)\); Parket al\.\([2023](https://arxiv.org/html/2608.18085#bib.bib6)\)\.

Other studies move from measuring personality to actively inducing or editing it, for example through prompt design, persona conditioning, or model\-level control mechanismsMaoet al\.\([2023](https://arxiv.org/html/2608.18085#bib.bib12)\); Lotfiet al\.\([2023](https://arxiv.org/html/2608.18085#bib.bib1)\)\. However, these studies mostly evaluate personality realization in open\-ended or non\-task settings\. They therefore show that personality can be represented and manipulated in LLM outputs, but do not establish how personality conditioning interacts with schema constraints, API decisions, or multi\-turn task completion\.

### 2\.3Personalized Dialogue and User Modeling

Personalized dialogue research studies how systems adapt to user characteristics, preferences, or inferred conversational tendencies\. Recent surveys frame this direction as part of LLM personalization, where models condition on user\-specific context, such as facts, style, persona descriptions, or trait\-like profiles, to make conversations more relevant and coherent over timeTsenget al\.\([2024](https://arxiv.org/html/2608.18085#bib.bib22)\); Zhanget al\.\([2018](https://arxiv.org/html/2608.18085#bib.bib21)\); Lotfiet al\.\([2023](https://arxiv.org/html/2608.18085#bib.bib1)\)\.

Recent LLM\-based work has also explored persona\-aware dialogue generation, where user profiles or dialogue history are used to improve persona consistency across interactionsLiuet al\.\([2025](https://arxiv.org/html/2608.18085#bib.bib24)\)\. This line of work is related to our focus on personality\-aware interaction, but it is not designed for schema\-guided task\-oriented dialogue, where systems must satisfy slot constraints, issue valid API calls, and complete user goals\. In parallel, LLM\-based TOD user simulation uses prompted LLMs to generate goal\-driven user behavior for training or evaluating dialogue systemsAlgherairy and Ahmed \([2025](https://arxiv.org/html/2608.18085#bib.bib23)\); Davidsonet al\.\([2023](https://arxiv.org/html/2608.18085#bib.bib25)\)\.

Taken together, prior work has studied TOD and personality largely separately\. TOD research emphasizes structure, grounding, and task validity, whereas personality research emphasizes trait expression, persona consistency, and controllability\. Our framework connects these directions by conditioning the simulated user on Big Five traits and evaluating how personality access affects schema\-grounded system behavior, pairwise user satisfaction, and trait realization across multiple models and evaluation metrics\.

## 3Methods and Experimental Setup

We build a controlled evaluation framework for personality\-aware task\-oriented dialogue using the Schema\-Guided Dialogue \(SGD\) datasetRastogiet al\.\([2020](https://arxiv.org/html/2608.18085#bib.bib13)\)\. The framework simulates interactions between an LLM\-based user agent and an LLM\-based system agent under matched task settings\. Across conditions, the task schema, domain, demonstration example, and dialogue\-generation rules are held fixed; the main experimental variable is the system agent’s access to personality information, following theOracle,Try, andNeutralsettings introduced in Section[1](https://arxiv.org/html/2608.18085#S1)\.

### 3\.1Task\-Oriented Dialogue Framework

Each dialogue is grounded in an SGD task schema that specifies the domain, intents, slots, and API structure required to complete the task\. Both agents operate in a one\-shot prompting setup without fine\-tuning\. At each turn, generation is conditioned on an in\-domain demonstration, the relevant schema and API rules, operational instructions, and the dialogue history up to turntt\.

The system agent produces either a natural\-language response or an API call\. Following prior workMosharrofet al\.\([2025](https://arxiv.org/html/2608.18085#bib.bib7)\), we represent an API call as

APICall=\(I,\{\(si,vi\)\}\),\\text\{APICall\}=\\big\(I,\\\{\(s\_\{i\},v\_\{i\}\)\\\}\\big\),whereIIis the domain’s intent and\(si,vi\)\(s\_\{i\},v\_\{i\}\)are slot–value assignments\. This representation lets us evaluate whether the dialogue remains grounded in the task schema and whether the system completes the intended search or reservation goal with valid slot values\.

### 3\.2Personality Trait Specification

We model user personality using the Big Five framework, which captures broad behavioral variation across five dimensionsDigman \([1990](https://arxiv.org/html/2608.18085#bib.bib15)\); Goldberg \([2013](https://arxiv.org/html/2608.18085#bib.bib16)\); Wiggins \([1996](https://arxiv.org/html/2608.18085#bib.bib17)\); De Raad \([2000](https://arxiv.org/html/2608.18085#bib.bib18)\); Jianget al\.\([2024](https://arxiv.org/html/2608.18085#bib.bib19)\)\. To evaluate both canonical Big Five traits and opposite\-pole variants, each dialogue is assigned one target persona from the following ten traits: Extraversion, introversive, Agreeableness, disagreeable, Conscientiousness, unconscientious, Neuroticism, stable, Openness, and closed\.

In each run, the user agent is assigned exactly one target trait\. The user is instructed to express the trait implicitly through natural dialogue behavior rather than explicitly naming the trait\. This design allows us to test whether personality can be realized in task\-oriented interaction while the system still satisfies schema and API constraints\.

### 3\.3Prompting and Personality\-Access Conditions

Both agents are prompted with the same general structure: task instructions, domain schema and API specification, an in\-domain TOD demonstration dialogue, operational rules, and the dialogue history\. Personality information is added only when required by the experimental condition\.

The user agent receives the target trait and a short trait description, which guide how the user asks questions, reacts to recommendations, handles uncertainty, and decides whether to proceed with a booking\. For example, an open user may ask to explore additional options, a neurotic user may seek reassurance, and a disagreeable user may challenge suggestions or request alternatives more often\.

The system agent is always instructed to remain task\-valid, grounded, and schema\-compliant\. In personality\-aware conditions, it may adapt its interaction strategy to the user’s personality by adjusting the level of detail, reassurance, recommendation style, or clarification strategy\. This setup allows us to test whether personality access improves system quality and user satisfaction, and whether such adaptation introduces grounding risks\.

### 3\.4Domains, Models, and Dialogue Generation

We conduct experiments in the Restaurant and Hotel domains of the Schema\-Guided Dialogue \(SGD\) datasetRastogiet al\.\([2020](https://arxiv.org/html/2608.18085#bib.bib13)\)\. Both domains involve schema\-guided search and reservation tasks, but differ in how strongly user preferences shape the interaction\. Restaurant dialogues often require more preference negotiation, whereas Hotel dialogues depend more directly on constraints such as location, star rating, and number of rooms\.

For each combination of domain, model, personality\-access condition, and target trait, we generate a fixed set of dialogues using the same schema constraints and prompting protocol\. The user always initiates the interaction\. Dialogues stop when the scripted completion phrase is reached after task completion, after 70 turns, or after 5 API calls\.

We use the same experimental structure across GPT\-4o, Qwen3\-Next\-80B, and Gemini 2\.0 Flash, with minor model\-specific prompt\-formatting adjustments to ensure reliable generation and valid API\-call formatting\. A representative Restaurant\-domainOraclesystem\-agent prompt template is provided in Appendix[B\.1](https://arxiv.org/html/2608.18085#A2.SS1)\. Full experimental settings, including generation parameters, model access routes, computational resources, and preprocessing details, are provided in Appendix[D](https://arxiv.org/html/2608.18085#A4)\.

### 3\.5Evaluation Metrics

We evaluate the generated dialogues using four groups of complementary metrics\.

1. 1\.Task outcomes\.Dialogue Completionmeasures whether the interaction reaches a functional terminal state and ends naturally\.Task Successmeasures whether the intended schema\-constrained task is completed correctly, including the appropriate final intent and required slot values\. We report these metrics separately because a dialogue can end naturally without completing the task, or complete the task without reaching the scripted terminal phrase\.
2. 2\.Personality realization\.We measure dialogue\-level trait realization using the TOD Trait Score, which captures how clearly the user agent expresses the intended trait in generated task\-oriented dialogues\. Separately, we use a 44\-item Big Five Inventory \(BFI\) questionnaire probeJohn and Srivastava \([1999](https://arxiv.org/html/2608.18085#bib.bib26)\)to validate whether each trait prompt aligns with the intended Big Five direction\. The BFI score is not computed from generated TOD dialogues; it serves as a controlled prompt\-level check, while Avg\. Intended Score measures trait realization within the generated conversations\.
3. 3\.System quality\.We use an LLM judge to score three system\-quality dimensions\.Constraint Satisfactionmeasures whether the final recommendation, result, or booking correctly fulfills the user’s active task constraints\.Inform Ratemeasures whether the system explicitly provides correct requested information or result details that are supported by the dialogue and API/search results\.Truthfulnessmeasures whether the system avoids unsupported or fabricated claims and remains grounded in the dialogue and available API/search results\. All three metrics are scored on a 1–5 scale, where higher is better\.
4. 4\.Pairwise user satisfaction\.To assess whether personality\-aware interactions are perceived as more satisfying than neutral TOD interactions, we compare paired dialogues generated under different personality\-access conditions while matching domain, model, and target trait\. The evaluator selects which dialogue would be more satisfying from the user’s perspective\. We report pairwise win rates forOracleversusNeutral,TryversusNeutral, andTryversusOracle\.

For LLM\-based evaluation of system quality and dialogue\-level user personality realization, we use Gemma\-4\-31B as the judge model\. The same judge is used across conditions to keep evaluation consistent\. For the BFI probe, we evaluate trait alignment separately using the same three LLMs used for dialogue generation: GPT\-4o, Qwen3\-Next\-80B, and Gemini 2\.0 Flash\. The evaluator prompts used for system\-quality, pairwise\-satisfaction, personality\-realization, and BFI judgments are provided in Appendix[B](https://arxiv.org/html/2608.18085#A2)\.

## 4Results and Analysis

We organize the results around the three research questions introduced in Section[1](https://arxiv.org/html/2608.18085#S1)\. Across the Hotel and Restaurant domains, we analyze how the three system personality\-access conditions—Oracle,Try, andNeutral—affect task performance, dialogue quality, and personality\-related outcomes\.

### 4\.1RQ1: Can LLMs jointly express personality and maintain schema\-guided task performance?

To answer RQ1, Table[1](https://arxiv.org/html/2608.18085#S4.T1)summarizes task and personality\-expression outcomes averaged across domains, models, and system personality\-access conditions\. The framework maintains strong schema\-guided performance overall, reaching 89\.94% Dialogue Completion and 85\.27% Task Success\. Thus, personality\-conditioned user behavior does not generally prevent dialogues from reaching a coherent endpoint or completing the underlying API\-grounded task\.

Personality expression is also evident, but varies by trait\. The User Trait Score, which measures the extent to which the user agent expresses the intended trait in generated task\-oriented dialogues, averages 3\.620\. It shows strong dialogue\-level realization for Extraversion, Agreeableness, Neuroticism, Openness, and disagreeable, but much weaker realization forintroversiveandclosed\. Separately, the BFI questionnaire probe yields a high average BFI Score of 4\.888, indicating that the trait specifications themselves align well with their intended Big Five directions\. This gap between strong BFI alignment and less uniform TOD trait realization suggests that some traits are easier to specify than to express consistently in grounded, multi\-turn task\-oriented dialogue\.

Overall, these results answerRQ1: training\-free LLM agents can jointly support personality\-aware interaction and schema\-guided task completion\.

Finding 1: Personality\-aware TOD is feasible, but dialogue\-level trait realization is unevenThe framework achieves strong overall task performance while expressing user traits in dialogue\. However, personality realization is uneven: several traits are strongly manifested in TOD interactions, whereasintroversiveandclosedremain difficult to realize reliably, despite strong alignment in the separate BFI probe\.

Table 1:Trait\-level task and personality realization outcomes averaged across domains, models, and the three system personality\-access conditions\. Dialogue Completion and Task Success summarize schema\-guided task performance\. User Trait Score measures dialogue\-level realization of the intended user trait in generated task\-oriented dialogues, while BFI Score measures prompt\-level semantic alignment with the intended Big Five direction using a separate questionnaire probe\.
### 4\.2RQ2: How does personality access affect system quality and pairwise user satisfaction?

To answer RQ2, Table[2](https://arxiv.org/html/2608.18085#S4.T2)compares system\-quality metrics and pairwise user\-satisfaction preferences across traits, averaged over domains and LLMs\. Overall, personality\-aware conditions improve system quality relative to the non\-personalized baseline\.Tryachieves the highest average Constraint Satisfaction \(4\.646\) and Inform Rate \(4\.849\), followed closely byOracle\(4\.598 and 4\.825\), whileNeutralis lower on both metrics \(4\.502 and 4\.741\)\. This suggests that access to personality information, especially when inferred from dialogue cues, helps the system respond more effectively to the user’s needs while preserving task\-relevant content\.

We test the reliability of these differences using Wilcoxon signed\-rank tests on dialogue\-level score differences, with Holm correction across the nine condition–metric comparisons\. The design is fully balanced: 3 models×\\times2 domains×\\times10 traits×\\times3 conditions×\\times50 dialogues, yielding 9,000 total dialogue\-level observations\. For each pairwise condition comparison, scores are paired within model–domain–trait cells, resulting in50×3×2×10=3,00050\\times 3\\times 2\\times 10=3\{,\}000paired differences\. Personality\-aware conditions significantly improve Constraint Satisfaction overNeutral\(Try:W=95682W\{=\}95682,p<\.001p\{<\}\.001;Oracle:W=114496W\{=\}114496,p<\.001p\{<\}\.001\) and Inform Rate \(Try:W=35335W\{=\}35335,p<\.001p\{<\}\.001;Oracle:W=44284W\{=\}44284,p<\.001p\{<\}\.001\)\. However, both adaptive conditions significantly reduce Truthfulness relative toNeutral\(Try:W=68192W\{=\}68192,p<\.001p\{<\}\.001;Oracle:W=61759W\{=\}61759,p<\.001p\{<\}\.001\), confirming a personalization–grounding trade\-off\. Comparing the two adaptive conditions,TryandOracledo not differ significantly on Constraint Satisfaction or Inform Rate, butTrypreserves significantly higher Truthfulness thanOracle\(W=119937W\{=\}119937,p<\.001p\{<\}\.001\)\. This trade\-off is especially visible for traits such as Extraversion and Openness, where explicit personality conditioning lowers Truthfulness despite improving or maintaining other quality dimensions\.

The pairwise satisfaction results show that personality access is generally preferred by evaluators\.Oracleis favored overNeutralat an average rate of 59\.63% versus 39\.87%, whileTryis preferred overNeutraleven more strongly, at 63\.40% versus 35\.20%\. Directly comparing the two adaptive strategies,Tryis preferred overOracleoverall \(56\.10% vs\. 43\.00%\), indicating that cue\-based adaptation yields the strongest aggregate user\-facing benefit\.

Trait\-level patterns further clarify this result\.Oracleperforms especially well againstNeutralfor Openness, Extraversion, and Agreeableness, suggesting that explicit trait knowledge can be helpful when the target style is socially clear and easy to support\. In contrast,Tryis preferred overOraclefor introversive, disagreeable, closed, and unconscientious users\. The disagreeable trait is particularly informative\. In the aggregated table,OracleunderperformsNeutralin pairwise satisfaction for disagreeable users \(28\.33% vs\. 71\.67%\), whileTryremains closer to the baseline \(47\.00% vs\. 53\.00%\) and is strongly preferred overOracle\(80\.66% vs\. 19\.34%\)\. This decline is largely driven by Qwen3 and Gemini, which, in the disagreeableOraclesetting, often mirror the customer’s hostile and antagonistic tone rather than adapting to it constructively; this failure is less pronounced underTry\. GPT\-4o behaves differently: in the Restaurant domain, it handles disagreeable users effectively, withOraclepreferred overNeutralby 70% to 30%, andTrypreferred overNeutralby 64% to 36%\. Detailed model\- and domain\-level results are reported in Table[9](https://arxiv.org/html/2608.18085#A1.T9)in Appendix[A\.4](https://arxiv.org/html/2608.18085#A1.SS4)\.

Taken together, these results answerRQ2: personality access improves system quality and pairwise satisfaction, but cue\-based adaptation provides the most reliable balance between personalization and grounding\.

Finding 2: Personality access improves system quality and satisfaction, but cue\-based adaptation inTryis more robustPersonality\-aware conditions outperform theNeutralbaseline on Constraint Satisfaction, Inform Rate, and pairwise user satisfaction, althoughNeutralremains the most truthful condition\. Among adaptive strategies,Tryprovides the strongest overall balance: it achieves the highest average system\-quality scores and is preferred over bothNeutralandOraclein pairwise satisfaction\. ExplicitOracleconditioning is beneficial for several traits, but it can over\-amplify socially difficult styles, as seen for disagreeable users in Qwen3 and Gemini\.

Table 2:Trait\-level system\-quality scores and pairwise user\-satisfaction preferences averaged across domains and LLMs\. N, T, and O denote Neutral, Try, and Oracle, respectively\. CS, IR, and TR denote Constraint Satisfaction, Inform Rate, and Truthfulness, each reported on a 1–5 scale, where higher is better\. Pairwise satisfaction cells report win\-rate percentages for the first condition over the second; for example, O/N = 62/38 indicates that Oracle is preferred in 62% of comparisons and Neutral in 38%\. Values may sum to less than 100 due to ties\.
### 4\.3RQ3: Do satisfaction gains depend on user trait realization?

Finally, we examine whether satisfaction gains depend on dialogue\-level trait realization\. We correlate Avg\. User Trait Score with pairwise satisfaction gainsΔO​\-​N\\Delta\_\{O\\text\{\-\}N\}andΔT​\-​N\\Delta\_\{T\\text\{\-\}N\}using Spearman rank correlation which is shown in Figure[2](https://arxiv.org/html/2608.18085#S4.F2)\. As shown in Table[3](https://arxiv.org/html/2608.18085#S4.T3),Oraclegains increase significantly with stronger trait realization:ΔO​\-​N\\Delta\_\{O\\text\{\-\}N\}correlates with Avg\. Trait Score atρ=\+0\.685\\rho\{=\}\{\+\}0\.685\(p=\.029p\{=\}\.029,n=10n\{=\}10\)\. Strongly realized traits such as Extraversion, Openness, and Agreeableness show largerOracle\-over\-Neutralgains, whereas weakly realized traits such asintroversiveandclosedshow smaller or negative gains\. This suggests that explicit trait conditioning is most useful when the user’s dialogue provides clear behavioral evidence for the supplied trait label\.

This pattern also appears beyond the trait\-level aggregate\. At the model×\\timesdomain×\\timestrait cell level, the association remains positive and significant, though more moderate \(ρ=\+0\.421\\rho\{=\}\{\+\}0\.421,p=\.001p\{=\}\.001,n=60n\{=\}60, 95% CI \[0\.19, 0\.61\]\)\. The effect is strongest for GPT\-4o \(ρ=\+0\.682\\rho\{=\}\{\+\}0\.682,p<\.001p\{<\}\.001\) and remains positive for Gemini \(ρ=\+0\.453\\rho\{=\}\{\+\}0\.453,p=\.044p\{=\}\.044\), indicating that the realization–satisfaction relationship is partly model\-dependent\.

By contrast,Trygains do not meaningfully depend on trait realization\. At the trait level,ΔT​\-​N\\Delta\_\{T\\text\{\-\}N\}is not significantly correlated with Avg\. Intended Score \(ρ=−0\.103\\rho\{=\}\{\-\}0\.103,p=\.777p\{=\}\.777\), and the same holds at the cell level \(ρ=−0\.157\\rho\{=\}\{\-\}0\.157,p=\.232p\{=\}\.232\)\. BecauseTryadapts to cues observed during the dialogue rather than relying on a predefined trait label, its benefits are less sensitive to whether the intended trait is strongly expressed\.

Together, these results answerRQ3: satisfaction gains from explicit personality conditioning depend on whether the target trait is clearly realized, whereas cue\-based adaptation is less sensitive to realization strength\.

Finding 3: Explicit personality conditioning depends on clear trait realizationOraclegains increase with stronger dialogue\-level trait realization, whileTrygains are not meaningfully tied to realization strength\. Cue\-based adaptation inTryprovides more stable gains across traits\.

Table 3:Spearman correlations between Avg\. User Trait Score and pairwise satisfaction deltas at two levels of analysis\. Trait\-level uses one point per trait \(n=10n\{=\}10\), averaged across models and domains\. Cell\-level uses model×\\timesdomain×\\timestrait cells \(n=60n\{=\}60, Tables 8–9\)\.p∗<\.05\{\}^\{\*\}p\{<\}\.05;p∗⁣∗∗<\.001\{\}^\{\*\*\*\}p\{<\}\.001\.![Refer to caption](https://arxiv.org/html/2608.18085v1/latex/figures/Spearman_Test/fig_score_only.png)Figure 2:Spearman correlation between Avg\. User Intended Score and pairwise satisfaction deltas across traits \(n=10n\{=\}10\)\. Oracle gains increase with trait realization \(ρ=\+0\.685\\rho\{=\}\{\+\}0\.685,p=\.029p\{=\}\.029\), whereas Try gains do not \(ρ=−0\.103\\rho\{=\}\{\-\}0\.103,p=\.777p\{=\}\.777\)\. Point size indicates cross\-dataset SD; disagreeable reflects a Qwen3/Gemini mirroring artifact \(Appendix[C\.3](https://arxiv.org/html/2608.18085#A3.SS3)\)\.

## 5Conclusion

We studied whether LLMs can express personality in goal\-directed task\-oriented dialogue without weakening task completion, and whether adapting to the user’s personality improves interaction quality\. To study this, we built a controlled, training\-free framework that simulates interactions between a personality\-conditioned user agent and a task\-oriented system agent across three LLMs, two SGD domains, and ten Big Five traits and opposite\-pole variants\.

Our results show that LLM agents can express personality while maintaining strong task performance, but trait realization is uneven: some traits are expressed clearly, whereas traits such asclosedandintroversiveare much harder to realize in multi\-turn TOD\. We also find that personality adaptation improves Constraint Satisfaction, Inform Rate, and pairwise user satisfaction, but reduces Truthfulness\. Wilcoxon signed\-rank tests confirm the reliability of these differences, revealing a clear personalization–grounding trade\-off\.

Finally, the benefit of personality adaptation depends on how the system obtains user personality information\.Oraclegains increase when the target trait is strongly expressed, suggesting that explicit personality labels help most when they match observable user behavior\. However, explicit conditioning can also over\-amplify difficult interaction styles\. In contrast,Tryis less sensitive to realization strength and provides the most reliable overall balance\. Overall, cue\-based adaptation inTryoffers a more robust route to personality\-aware TOD without fine\-tuning\. Future work should validate these findings with human users and explore hybrid strategies that combine explicit trait information with dialogue cues\.

## 6Limitations

This work has several limitations\. First, we study personality\-aware task\-oriented dialogue through simulated interactions between LLM\-based user and system agents\. This controlled design allows us to isolate the effect of personality access under matched task settings, but future work should validate the findings with real users and human\-centered interaction studies\.

Second, our experiments focus on the Hotel and Restaurant domains of the Schema\-Guided Dialogue dataset\. These domains provide structured search and reservation tasks with clear API grounding, but they do not cover all types of task\-oriented dialogue\. Extending the framework to domains with different risk profiles, such as education, healthcare, or customer support, would help test whether the same personalization–grounding trade\-off holds more broadly\.

Third, we assign one target personality trait per dialogue using the Big Five traits and their opposite poles\. This design supports controlled trait\-level comparison, but real users may express multiple traits simultaneously or shift their interaction style across turns\. Future work could study mixed\-trait personas and dynamic personality expression\.

Fourth, our evaluation uses LLM\-based judges for system quality, pairwise satisfaction, and trait realization\. These judges enable scalable comparison across many models, traits, and domains, but they may not fully replace human preference judgments\. Additional human evaluation would strengthen the interpretation of user satisfaction and personality alignment\.

Finally, our study focuses on prompting\-based personality adaptation without fine\-tuning\. Different model families, decoding settings, prompt designs, or tool\-use implementations may lead to different outcomes\. Therefore, our findings should be interpreted as evidence for training\-free personality\-aware prompting in controlled schema\-guided dialogue settings\.

## References

- Prompting large language models for user simulation in task\-oriented dialogue systems\.Computer Speech & Language89,pp\. 101697\.Cited by:[§2\.3](https://arxiv.org/html/2608.18085#S2.SS3.p2.1)\.
- S\. Davidson, S\. Romeo, R\. Shu, J\. Gung, A\. Gupta, S\. Mansour, and Y\. Zhang \(2023\)User simulation with large language models for evaluating task\-oriented dialogue\.arXiv preprint arXiv:2309\.13233\.Cited by:[§2\.3](https://arxiv.org/html/2608.18085#S2.SS3.p2.1)\.
- B\. De Raad \(2000\)The Big Five personality factors: the psycholexical approach to personality\.\.Hogrefe & Huber Publishers\.Cited by:[§3\.2](https://arxiv.org/html/2608.18085#S3.SS2.p1.1)\.
- J\. M\. Digman \(1990\)Personality structure: emergence of the five\-factor model\.Annual Review of Psychology41\(1\),pp\. 417–440\.Cited by:[§3\.2](https://arxiv.org/html/2608.18085#S3.SS2.p1.1)\.
- M\. Eric, R\. Goel, S\. Paul, A\. Sethi, S\. Agarwal, S\. Gao, A\. Kumar, A\. Goyal, P\. Ku, and D\. Hakkani\-Tur \(2020\)MultiWOZ 2\.1: a consolidated multi\-domain dialogue dataset with state corrections and state tracking baselines\.InProceedings of the Twelfth Language Resources and Evaluation Conference,pp\. 422–428\.Cited by:[§2\.1](https://arxiv.org/html/2608.18085#S2.SS1.p1.1)\.
- L\. R\. Goldberg \(2013\)An alternative “description of personality”: the Big\-Five factor structure\.InPersonality and Personality Disorders,pp\. 34–47\.Cited by:[§1](https://arxiv.org/html/2608.18085#S1.p2.1),[§3\.2](https://arxiv.org/html/2608.18085#S3.SS2.p1.1)\.
- T\. Han, X\. Liu, R\. Takanabu, Y\. Lian, C\. Huang, D\. Wan, W\. Peng, and M\. Huang \(2021\)MultiWOZ 2\.3: a multi\-domain task\-oriented dialogue dataset enhanced with annotation corrections and co\-reference annotation\.InCCF International Conference on Natural Language Processing and Chinese Computing,pp\. 206–218\.Cited by:[§2\.1](https://arxiv.org/html/2608.18085#S2.SS1.p1.1)\.
- G\. Jiang, M\. Xu, S\. Zhu, W\. Han, C\. Zhang, and Y\. Zhu \(2022\)MPI: evaluating and inducing personality in pre\-trained language models\.arXiv preprint arXiv:2206\.07550\.Cited by:[§1](https://arxiv.org/html/2608.18085#S1.p2.1),[§2\.2](https://arxiv.org/html/2608.18085#S2.SS2.p1.1)\.
- H\. Jiang, X\. Zhang, X\. Cao, C\. Breazeal, D\. Roy, and J\. Kabbara \(2024\)PersonaLLM: investigating the ability of large language models to express personality traits\.InFindings of the Association for Computational Linguistics: NAACL 2024,pp\. 3605–3627\.Cited by:[§1](https://arxiv.org/html/2608.18085#S1.p2.1),[§2\.2](https://arxiv.org/html/2608.18085#S2.SS2.p1.1),[§3\.2](https://arxiv.org/html/2608.18085#S3.SS2.p1.1)\.
- O\. P\. John and S\. Srivastava \(1999\)The big five trait taxonomy: history, measurement, and theoretical perspectives\.InHandbook of Personality: Theory and Research,L\. A\. Pervin and O\. P\. John \(Eds\.\),pp\. 102–138\.Cited by:[item 2](https://arxiv.org/html/2608.18085#S3.I1.i2.p1.1)\.
- X\. Li, Y\. Li, L\. Liu, L\. Bing, and S\. Joty \(2022\)Is gpt\-3 a psychopath? evaluating large language models from a psychological perspective\.arXiv preprint arXiv:2212\.1052910\.Cited by:[§1](https://arxiv.org/html/2608.18085#S1.p2.1),[§2\.2](https://arxiv.org/html/2608.18085#S2.SS2.p1.1)\.
- D\. Liu, Z\. Wu, D\. Song, and H\. Huang \(2025\)A persona\-aware llm\-enhanced framework for multi\-session personalized dialogue generation\.InFindings of the Association for Computational Linguistics: ACL 2025,pp\. 103–123\.Cited by:[§2\.3](https://arxiv.org/html/2608.18085#S2.SS3.p2.1)\.
- E\. Lotfi, M\. De Bruyn, J\. Buhmann, and W\. Daelemans \(2023\)PersonalityChat: conversation distillation for personalized dialog modeling with facts and traits\.InProceedings of the Third Workshop on Natural Language Generation, Evaluation, and Metrics \(GEM\),pp\. 353–371\.Cited by:[§2\.2](https://arxiv.org/html/2608.18085#S2.SS2.p2.1),[§2\.3](https://arxiv.org/html/2608.18085#S2.SS3.p1.1)\.
- S\. Mao, N\. Zhang, X\. Wang, M\. Wang, Y\. Yao, Y\. Jiang, P\. Xie, F\. Huang, and H\. Chen \(2023\)Editing personality for LLMs\.CoRR\.Cited by:[§1](https://arxiv.org/html/2608.18085#S1.p2.1),[§2\.2](https://arxiv.org/html/2608.18085#S2.SS2.p2.1)\.
- A\. Mosharrof, M\. Fereidouni, and A\. Siddique \(2025\)Evaluating and enhancing out\-of\-domain generalization of task\-oriented dialog systems for task completion without turn\-level dialog annotations\.arXiv preprint arXiv:2502\.13310\.Cited by:[§1](https://arxiv.org/html/2608.18085#S1.p3.1),[§2\.1](https://arxiv.org/html/2608.18085#S2.SS1.p2.1),[§3\.1](https://arxiv.org/html/2608.18085#S3.SS1.p2.3)\.
- K\. Pan and Y\. Zeng \(2023\)Do LLMs possess a personality? making the MBTI test an amazing evaluation for large language models\.arXiv preprint arXiv:2307\.16180\.Cited by:[§1](https://arxiv.org/html/2608.18085#S1.p2.1),[§2\.2](https://arxiv.org/html/2608.18085#S2.SS2.p1.1)\.
- J\. S\. Park, J\. O’Brien, C\. J\. Cai, M\. R\. Morris, P\. Liang, and M\. S\. Bernstein \(2023\)Generative agents: interactive simulacra of human behavior\.InProceedings of the 36th annual acm symposium on user interface software and technology,pp\. 1–22\.Cited by:[§1](https://arxiv.org/html/2608.18085#S1.p1.1),[§2\.2](https://arxiv.org/html/2608.18085#S2.SS2.p1.1)\.
- A\. Rastogi, X\. Zang, S\. Sunkara, R\. Gupta, and P\. Khaitan \(2020\)Towards scalable multi\-domain conversational agents: the schema\-guided dialogue dataset\.InProceedings of the AAAI Conference on Artificial Intelligence,Vol\.34,pp\. 8689–8696\.Cited by:[§1](https://arxiv.org/html/2608.18085#S1.p3.1),[§1](https://arxiv.org/html/2608.18085#S1.p4.1),[§2\.1](https://arxiv.org/html/2608.18085#S2.SS1.p1.1),[§3\.4](https://arxiv.org/html/2608.18085#S3.SS4.p1.1),[§3](https://arxiv.org/html/2608.18085#S3.p1.1)\.
- G\. Serapio\-García, M\. Safdari, C\. Crepy, L\. Sun, S\. Fitz, M\. Abdulhai, A\. Faust, and M\. Matarić \(2023\)Personality traits in large language models\.arXiv preprint\.Cited by:[§1](https://arxiv.org/html/2608.18085#S1.p2.1),[§2\.2](https://arxiv.org/html/2608.18085#S2.SS2.p1.1)\.
- Y\. Tseng, Y\. Huang, T\. Hsiao, W\. Chen, C\. Huang, Y\. Meng, and Y\. Chen \(2024\)Two tales of persona in llms: a survey of role\-playing and personalization\.InFindings of the Association for Computational Linguistics: EMNLP 2024,pp\. 16612–16631\.Cited by:[§2\.3](https://arxiv.org/html/2608.18085#S2.SS3.p1.1)\.
- N\. Wang, Z\. Peng, H\. Que, J\. Liu, W\. Zhou, Y\. Wu, H\. Guo, R\. Gan, Z\. Ni, and J\. Yang \(2024\)RoleLLM: benchmarking, eliciting, and enhancing role\-playing abilities of large language models\.InFindings of the Association for Computational Linguistics: ACL 2024,pp\. 14743–14777\.Cited by:[§1](https://arxiv.org/html/2608.18085#S1.p1.1),[§2\.2](https://arxiv.org/html/2608.18085#S2.SS2.p1.1)\.
- J\. S\. Wiggins \(1996\)The five\-factor model of personality: theoretical perspectives\.Guilford Press\.Cited by:[§3\.2](https://arxiv.org/html/2608.18085#S3.SS2.p1.1)\.
- X\. Zang, A\. Rastogi, S\. Sunkara, R\. Gupta, J\. Zhang, and J\. Chen \(2020\)MultiWOZ 2\.2: a dialogue dataset with additional annotation corrections and state tracking baselines\.arXiv preprint arXiv:2007\.12720\.Cited by:[§1](https://arxiv.org/html/2608.18085#S1.p3.1),[§2\.1](https://arxiv.org/html/2608.18085#S2.SS1.p1.1)\.
- S\. Zhang, E\. Dinan, J\. Urbanek, A\. Szlam, D\. Kiela, and J\. Weston \(2018\)Personalizing dialogue agents: i have a dog, do you have pets too?\.InProceedings of the 56th Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 2204–2213\.Cited by:[§2\.3](https://arxiv.org/html/2608.18085#S2.SS3.p1.1)\.

## Appendix ASupplementary Experimental Results

### A\.1Task Success Rates

Table[4](https://arxiv.org/html/2608.18085#A1.T4)reports task success rates across models, domains, and personality\-access conditions\.

Table 4:Task success rates \(%\) across models, domains, and personality\-access conditions\. In the Hotel domain, Neutral achieves the highest cross\-model average task success, whereas in the Restaurant domain, Oracle performs best, followed by Try\.
### A\.2Dialogue Completion Rates

Table[5](https://arxiv.org/html/2608.18085#A1.T5)reports dialogue\-completion rates across all models, traits, domains, and personality\-access conditions\.

Table 5:Dialogue\-completion rates \(%\) for GPT\-4o, Qwen3, and Gemini across personality traits in the Hotel and Restaurant domains\. N, T, and O denote Neutral, Try, and Oracle, respectively\. The average row reports the mean dialogue\-completion rate across the ten personality traits for each model and domain\.
### A\.3System\-Quality Scores

Tables[6](https://arxiv.org/html/2608.18085#A1.T6)and[7](https://arxiv.org/html/2608.18085#A1.T7)report trait\-level system\-quality scores for the Restaurant and Hotel domains, respectively\.

Table 6:Restaurant\-domain system\-evaluation results for GPT\-4o, Qwen3, and Gemini across personality traits under Neutral \(N\), Try \(T\), and Oracle \(O\) conditions\. The table reports Constraint Satisfaction, Truthfulness, and Inform Rate; the average row reports the mean across the ten personality traits for each model\.Table 7:Hotel\-domain system\-evaluation results for GPT\-4o, Qwen3, and Gemini across personality traits under Neutral \(N\), Try \(T\), and Oracle \(O\) conditions\. The table reports Constraint Satisfaction, Truthfulness, and Inform Rate; the average row reports the mean across the ten personality traits for each model\.
### A\.4Pairwise Satisfaction and Trait Realization

Tables[8](https://arxiv.org/html/2608.18085#A1.T8)and[9](https://arxiv.org/html/2608.18085#A1.T9)report pairwise user\-satisfaction deltas together with averaged trait\-realization metrics for the Hotel and Restaurant domains\.

Table 8:Hotel\-domain pairwise user\-satisfaction win rates with averaged personality\-realization metrics for GPT\-4o, Qwen3, and Gemini\. O, T, and N denote Oracle, Try, and Neutral, respectively\. Each comparison cell reports the win rate of the first condition over the second condition; for example, O/N = 36/58 means Oracle is preferred in 36% of paired comparisons and Neutral in 58%\. Percentages may sum to less than 100 when ties occur\. BFI Score is the intended\-direction score from the BFI probe for the target trait\. Avg\. User Trait Score and Avg\. Rank are averaged across Oracle, Try, and Neutral cases\. Lower Avg\. Rank indicates stronger realization of the intended trait\.Table 9:Restaurant\-domain pairwise user\-satisfaction win rates with averaged personality\-realization metrics for GPT\-4o, Qwen3, and Gemini\. O, T, and N denote Oracle, Try, and Neutral, respectively\. Each comparison cell reports the win rate of the first condition over the second condition, e\.g\., O/N = 62/38 means Oracle is preferred in 62% of paired comparisons and Neutral in 38%\. Percentages may sum to less than 100 when ties occur\. BFI Score is the intended\-direction score from the BFI probe for the target trait\. Avg\. User Trait Score and Avg\. Rank are averaged across Oracle, Try, and Neutral cases\. Lower Avg\. Rank indicates stronger realization of the intended trait\.
### A\.5Representative Model\- and Domain\-Specific Score\-Delta Figure

Figure[3](https://arxiv.org/html/2608.18085#A1.F3)provides a representative model\- and domain\-specific visualization of score deltas for GPT\-4o in the Restaurant domain\. We include this figure to illustrate how condition\-wise changes vary across traits for a single model\-domain setting\. The main paper reports aggregated cross\-model and cross\-domain results, while the appendix tables provide the full quantitative breakdown\.

![Refer to caption](https://arxiv.org/html/2608.18085v1/latex/figures/Restaurant/try_minus_neutral_delta_small_multiples.png)

![Refer to caption](https://arxiv.org/html/2608.18085v1/latex/figures/Restaurant/trait_minus_neutral_delta_small_multiples.png)

Figure 3:Representative GPT\-4o Restaurant\-domain score\-delta figures\. The top panel shows score changes fromNeutraltoTry, and the bottom panel shows score changes fromNeutraltoOracle\.

## Appendix BPrompts

### B\.1Dialogue\-Agent Prompt Templates

#### System agent prompt template: Restaurant Oracle condition\.

The following compact template illustrates the system\-agent instruction used for the Restaurant\-domainOraclecondition\. In this setting, the system is explicitly given the user’s target personality trait and its definition\.

Restaurant Oracle System\-Agent Prompt TemplateYou are an expert task\-oriented dialogue assistant for the Restaurant domain\. Your goal is to generate only the next SYSTEM response given the dialogue history\.You are given the user’s personality trait and its definition:\{personality\_definition\}Adapt your response style, tone, and interaction strategy to this personality while remaining helpful, grounded, and schema\-compliant\.Use the following Restaurant schema:FindRestaurantsRequired slots:city,cuisineOptional slots:price\_range,has\_live\_music,serves\_alcoholResult slots:restaurant\_name,serves\_alcohol,has\_live\_music,phone\_number,street\_address,price\_range,city,cuisineReserveRestaurantRequired slots:restaurant\_name,city,timeOptional slots:date,party\_sizeResult slots:restaurant\_name,date,time,serves\_alcohol,has\_live\_music,phone\_number,street\_address,party\_size,price\_range,city,cuisineRules:•Generate only the next SYSTEM utterance; do not generate USER turns\.•Use API calls instead of external knowledge when restaurant information is needed\.•Make API calls only after collecting the required slots\.•Format API calls as:APICall\(method=’intent\_name’,parameters=\{slot: value\}\)•Use only schema\-supported fields and database search results\.•Do not invent reviews, menus, ratings, or unsupported restaurant details\.•If search results are empty, explain the limitation and suggest a close alternative search\.•If the user requests a reservation, reserve the same restaurant previously discussed\.•Confirm key reservation slots before finalizing the reservation\.•If the user is open to any cuisine, select a cuisine from the allowed cuisine list based on the user’s personality and preferences\.•Stick to the city mentioned by the user; do not suggest nearby cities unless the user requests it\.Dialogue history:\{dialogue\_history\}Generate the next SYSTEM response\.

### B\.2User Trait Expression Evaluator Prompt

We used the following prompt to evaluate user trait realization\. The evaluator identifies the three most evident personality traits expressed by the USER across the full dialogue and assigns each selected trait a 1–5 expression score\.

User Trait Expression Evaluator PromptYou are an expert evaluator of personality expression in task\-oriented dialogue\.Your task is to identify the top three most evident personality traits expressed by the USER across the full dialogue and rate how strongly each trait is expressed, in ranked order\.Candidate traits:Extraversion, Agreeableness, Conscientiousness, Neuroticism, Openness, introversive, disagreeable, unconscientious, stable, and closed\.Rating scale:•1 = Very low expression of this trait•2 = Low expression of this trait•3 = Moderate or unclear expression•4 = High expression of this trait•5 = Very high expression of this traitInstructions:•Use the full dialogue rather than a single utterance\.•Judge only observable USER behavior, not SYSTEM behavior\.•Do not infer broad personality characteristics beyond what is clearly supported by the USER’s utterances\.•Return three distinct traits\.•Score each selected trait from 1 to 5\.•Keep reasons brief and evidence\-based\.•Return only valid JSON\.Dialogue:\{dialogue\}Return only valid JSON:``` { "top1_evident_trait": "", "top1_evident_trait_expression": { "score": 1, "reason": "" }, "top2_evident_trait": "", "top2_evident_trait_expression": { "score": 1, "reason": "" }, "top3_evident_trait": "", "top3_evident_trait_expression": { "score": 1, "reason": "" } } ```

### B\.3Pairwise User Satisfaction Evaluator Prompt

We used the following prompt to evaluate pairwise user satisfaction\. The evaluator compares two dialogues generated under the same domain, model, and intended user trait, and selects which system would feel more satisfying to the user\. Task success is excluded because it is evaluated separately\.

Pairwise User Satisfaction Evaluator PromptYou are evaluating pairwise user satisfaction in task\-oriented dialogues\.Compare two dialogues with the same intended user trait\. Choose which SYSTEM would feel more satisfying to the USER, considering the whole interaction\.Instructions:•Do not evaluate task success; it is measured separately\.•Do not penalize a system only because the API/search result is empty, unavailable, or limited, unless the system handles that limitation poorly\.•Focus on helpfulness, tone, responsiveness, patience, clarity, naturalness, and how well the system responds to the user’s needs and emotions\.•Also consider how well the system fits the user’s communication style\.•Do not choose only because one dialogue is shorter or more efficient, unless it clearly improves the user’s experience\.•Do not use the scripted closing sentence “Thanks a lot\! That’s all I need\.” as direct evidence of satisfaction, because it is a scripted ending\.•Choose “tie” only if both dialogues are similarly satisfying\.•Keep the reason brief and avoid quoting dialogue text directly\.•Return only valid JSON\.Dialogue A:\{DIALOGUE\_A\}Dialogue B:\{DIALOGUE\_B\}Return only valid JSON:\{ "winner": "A/B/tie", "reason": "" \}

### B\.4System Quality Evaluation Prompt

We used the following prompt to evaluate system\-level dialogue quality\. The evaluator scores each dialogue on four dimensions: Constraint Satisfaction, Inform Rate, and Dialogue Truthfulness\. Each criterion is rated on a 1–5 scale, with brief evidence\-based justification\.

System Quality Evaluator PromptYou are an expert evaluator of task\-oriented dialogues\.Read the full dialogue and rate each criterion on a scale of 1–5\.Evaluation Criteria:•Constraint Satisfaction Rate:Whether the final recommendation, result, or booking correctly satisfies the user’s active task constraints and requested outcome\.–1 = no correct or usable final result, or the result clearly violates important constraints–2 = weak match; only a small portion of important constraints is satisfied–3 = partially correct, but important constraints are missing, violated, or unclear–4 = mostly correct, with only minor omission or ambiguity–5 = fully correct, usable, and satisfies all active constraints•Inform Rate:Whether the system provides the requested information or result details correctly and with support from the dialogue or API/search results\.–1 = requested information is missing, clearly wrong, or unsupported–2 = only a small amount of correct information is provided, or several details are unsupported–3 = some correct information is provided, but important details are missing, incorrect, or unclear–4 = most requested information is correct and grounded, with only minor omission or ambiguity–5 = all requested information is clearly and correctly provided, and all details are supported•Dialogue Truthfulness Rate:Whether the system avoids unsupported claims, fabricated information, or misrepresentation of search/API results\.–1 = major hallucination, fabrication, or serious misrepresentation–2 = clear unsupported claim or multiple ungrounded details–3 = minor unsupported assumption or partially ungrounded statement–4 = mostly grounded, with at most a very small non\-material unsupported detail–5 = fully grounded; no meaningful hallucinationAvailable Data Constraints:•For Restaurant dialogues, available fields are: city name, cuisine type, live music, phone number, price range, restaurant name, alcohol service, and street address\.•For Hotel dialogues, available fields are: location, number of rooms, phone number, hotel name, price per night, smoking policy, star rating, and street address\.•Treat specific factual information beyond these fields as hallucinated unless it is explicitly supported in the dialogue\.Instructions:•Evaluate the entire dialogue, not only the final turn\.•Do not assume task success unless it is supported by the dialogue\.•Be strict and evidence\-based\.•Keep reasons brief and specific\.•Return only valid JSON, without markdown or extra commentary\.Dialogue:\{dialogue\}Return only valid JSON:\{ "Constraint Satisfaction Rate": \{ "score": 0, "reason": "" \}, "Inform Rate": \{ "score": 0, "reason": "" \}, "Dialogue Truthfulness Rate": \{ "score": 0, "reason": "" \}, \}

———————————————————

### B\.5BFI Questionnaire Probe

#### BFI questionnaire probe\.

We use a fixed questionnaire\-style probe to validate whether each trait prompt aligns with the intended Big Five direction\. The probe is separate from generated task\-oriented dialogues\.

BFI Trait Probe PromptAnswer the following BFI items as a realistic person matching the target trait\.Target trait:\{TARGET\_TRAIT\}Trait definition:\{TRAIT\_DEFINITION\}Use a 1–5 Likert scale: 1 = strongly disagree, 2 = disagree, 3 = neutral, 4 = agree, 5 = strongly agree\.BFI items:\{TRAIT\_SPECIFIC\_BFI\_ITEMS\}Return only valid JSON:``` { "target_trait": "{TARGET_TRAIT}", "responses": { "1": 1, "6": 1 } } ```

## Appendix CRealization–Satisfaction Correlation: Full Details

### C\.1Statistical Methodology

Each data point represents one trait×\\timesmodel×\\timesdomain cell, summarizing 50 pairwise comparisons per condition\. The satisfaction delta \(ΔO​\-​N\\Delta\_\{O\\text\{\-\}N\}orΔT​\-​N\\Delta\_\{T\\text\{\-\}N\}\) is the difference in win\-rates between the personality\-aware and Neutral conditions across those 50 binary outcomes\. Avg\. Intended Score is the mean expression score of the target trait averaged over all three conditions \(Oracle, Try, Neutral\) for that cell\.

We use Spearman rank correlation because \(i\) the unit of analysis is aggregated trait\-level data, not individual observations; \(ii\) the relationship is not assumed to be linear; and \(iii\) Spearman is robust to the bounded 1–5 scale and the non\-normal distribution of realization scores, where two traits \(introversive, closed\) have near\-zero scores while the remaining eight cluster near 4–5\. Results are reported at the trait\-averaged level \(n=10n\{=\}10\), which averages all six model–domain combinations per trait\. The pairwise satisfaction outcomes are binary \(win/loss/tie\); within\-cell significance can additionally be assessed using a binomial test againstH0:win rate=0\.5H\_\{0\}\\colon\\text\{win rate\}=0\.5\(n=50n\{=\}50pairs\), with 36 of 60 cells showing a significant deviation \(p<\.05p\{<\}\.05\)\.

### C\.2Per\-Trait Summary

Table[10](https://arxiv.org/html/2608.18085#A3.T10)reports per\-trait Avg\. Intended Score and average satisfaction deltas averaged across all six model–domain combinations\.

Table 10:Per\-trait Avg\. Intended Score and average pairwise satisfaction deltas \(pp\), averaged across six model–domain combinations\. Traits are sorted by Avg\. Intended Score in descending order\. The disagreeable trait is discussed separately in Appendix[C\.3](https://arxiv.org/html/2608.18085#A3.SS3)\.
### C\.3Disagreeable: System Mirroring in Qwen3 and Gemini

Under Oracle, Qwen3 and Gemini exhibit a system\-level artefact for thedisagreeabletrait: rather than adapting professionally to a hostile user, these models mirror the user’s antagonistic style, producing uncooperative responses and collapsing satisfaction \(ΔO​\-​N\\Delta\_\{O\\text\{\-\}N\}ranging from−56\-56to−92\-92pp across Hotel and Restaurant\)\. GPT\-4o does not exhibit this failure, maintaining cooperative, task\-focused behavior while acknowledging the user’s difficult style\. These four cells carry high realization scores \(Avg\. Score 4\.78–4\.96\) but anomalously negative deltas, suppressing the overall correlation\. Excluding them raises the result toρ=\+0\.557\\rho\{=\}\{\+\}0\.557\(p<\.001p\{<\}\.001,n=56n\{=\}56\), suggesting that the mirroring failure is a model\-specific prompt\-following limitation rather than a general property of personality conditioning\.

Note that absolute User Satisfaction scores for disagreeable under Oracle appear slightly higher than Neutral for Qwen3 and Gemini in some cells, which seemingly contradicts the pairwise preference for Neutral\. This discrepancy reflects a limitation of absolute rating scales near the floor: when both conditions produce poor interactions, a 1–5 score may lack the resolution to detect relative differences that pairwise comparison identifies more clearly\. The pairwise results are therefore the more informative evidence for the mirroring failure\.

## Appendix DReproducibility Details

Our experiments use a balanced design over three LLMs, two SGD domains \(Hotel and Restaurant\), ten personality traits, and three personality\-access conditions \(Neutral,Try, andOracle\)\. Each LLM generates 50 dialogues per domain–trait–condition cell, yielding 9,000 dialogue\-level observations\. For each pairwise condition comparison, scores are paired within model–domain–trait cells, giving50×3×2×10=3,00050\\times 3\\times 2\\times 10=3\{,\}000paired differences across both domains\. Trait\-level analyses average over models and domains, while cell\-level analyses use 60 model–domain–trait cells\. All dialogue\-generation runs use temperature=0\.7=0\.7, top\-p=1\.0p=1\.0, and a maximum output length of 250 tokens per LLM inference call\. Dialogues stop when the scripted completion phrase is reached after the goal is satisfied, after 70 turns, or after 5 API calls\. GPT\-4o, Qwen3\-Next\-80B, and Gemini 2\.0 Flash are used for dialogue generation\. Gemma\-4\-31B is used as the judge model for system quality, pairwise satisfaction, and dialogue\-level personality realization\. The BFI\-style questionnaire probe is evaluated using GPT\-4o, Qwen3\-Next\-80B, and Gemini 2\.0 Flash\. The main LLMs are accessed through hosted services rather than local training: GPT\-4o through OpenAI, Gemini 2\.0 Flash through OpenRouter, and Qwen3\-Next\-80B and Gemma\-4\-31B through Ollama Cloud\. Experiment management, dialogue parsing, metric aggregation, and statistical analysis are run on local computational resources with 32GB GPU memory\. CSV files of SGD domains are preprocessed usingpandas\. For each service domain, records are grouped by domain\-specific keys, inconsistent duplicate entries are canonicalized using the first record in each group\. For Restaurant results, records are grouped bystreet\_address,restaurant\_name, andcity, and duplicates are removed using fields such ascuisine,restaurant\_name,city,has\_live\_music,price\_range,serves\_alcohol,street\_address, andphone\_number\. This step does not introduce new service records or modify task schemas; it only removes redundant or inconsistent duplicate service entries\.