acceptodds
Desk rejected at ICLR 2027

Is Your Instruction Treated Fairly? Uncovering and Mitigating Expression Style-Induced Bias in LLM Agents

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

With the growing deployment of large language model-based agents (LLM agents), there is a heightened need to examine their fairness across diverse users. While prior research on fairness and bias has mainly centered on demographic attributes (e.g., gender and race), we identify a novel and practically significant source of disparity: the **expression style of user instructions**. Unlike task-relevant features such as instruction structure and information completeness, both of which are expected to affect task performance, expression style is inherently task-content-irrelevant and thus should not lead to performance disparities in an LLM agent. To systematically study this underexplored bias, we introduce the *Politeness-Arousal Style Framework* (), a two-dimensional taxonomy that characterizes user instructions along polite versus direct and calm versus emotional axes, yielding four distinct compound expression styles. Using PA-Style, we extensively evaluate various LLM agents, revealing substantial performance gaps across expression style conditions. To mitigate these disparities without sacrificing task success rate, we propose *Instruction Style Alignment* (), a lightweight and controllable approach that converts user instructions to a unified expression style. Through extensive experiments across diverse settings, we show that ISA effectively eliminates expression style-induced performance disparities while preserving task performance. Our code is available at [codes](https://anonymous.4open.science/r/isa_).

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM agents, Expression style-induced bias, fairness

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_b1hEOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents a study of how the expression style of user instructions affects the task performance of LLM agents, and frames the resulting performance gaps as an expression style-induced bias. The authors then propose the Politeness-Arousal Style Framework (PA-Style), which defines two style axes (polite vs. direct, calm vs. emotional) motivated by the Big Five model, and build a semantic-preserving stylization pipeline (LLM-based rewriting plus an LLM Semantic Comparator and Style Checker) on VitaBench to evaluate three agents under four compound styles. To reduce the gaps, they further propose Instruction Style Alignment (ISA), a lightweight BART-based rewriter trained with supervised warm-up and REINFORCE-style reinforcement learning, which converts user instructions into the style that the agent handles best and uses sampling-level fusion of two LoRA models to control the target style.

Critical strengths
  • The motivation is clearly explained, and the pilot study in Section 3.1 (Table 1) gives initial empirical evidence that changing only the style of an instruction can change agent behavior.

  • The authors release their code, which can support reproduction and future work.

Critical weaknesses
  • Mis-framed problem and limited novelty. The paper treats any style-induced performance difference as a fairness issue and claims to be the first to identify expression style as a bias dimension. However, fairness is defined with respect to sensitive or protected attributes[1], whereas the instruction is the main input of the agent. The phenomenon is better described as robustness to semantic-preserving input variations, and the effect of politeness and emotional expression on LLM performance has already been studied [2, 3]. Since the novel bias dimension is the central selling point of the paper, this substantially weakens its main contribution.

  • Insufficient evidence for both the measured bias and its mitigation. The stylized instructions are checked only by LLMs. Additionally, a single model plays the agent, the user simulator, and the evaluator (and apparently also the rewriter). The test set contains only 50 samples in a single-round setting, and ISA is evaluated on only one agent. Moreover, some results in Table 4 contradict the claim that ISA effectively eliminates the disparities (e.g., the Range on Pass^4 rises to 0.160). It is therefore unclear whether the reported disparities and their reduction reflect real agent behavior or artifacts of the evaluation pipeline.

[1] Machine Learning Testing: Survey, Landscapes and Horizons. TSE, 2020.

[2] Should We Respect LLMs? A Cross-Lingual Study on the Influence of Prompt Politeness on LLM Performance. SICon Workshop at ACL, 2024.

[3] Large Language Models Understand and Can Be Enhanced by Emotional Stimuli. arXiv:2307.11760, 2023.

Questions for authors
  1. Why should expression style be treated as a protected attribute rather than as an input variation that the agent should be robust to? Is there evidence that specific styles are strongly associated with specific user groups?
  2. Which model rewrites the instructions, and is it the same model as the agent and the evaluator? How do the authors rule out that the disparities come from the evaluator's own style preference or self-preference?
  3. What happens to instructions that still fail the Semantic Comparator or Style Checker after the maximum number of retries, and how many such cases are there? Has any human verification of semantic equivalence been performed?
  4. Why does ISA enlarge the Range on Pass^4 (0.160, +0.09), and why does the w/o Stage-2 variant achieve a higher Pass@4 Avg_styles (0.735) than the full ISA (0.650)? Which value of γ is used in the experiments?
  5. Please explain more clearly why the Big Five model is chosen as the theoretical basis and why only two of its five factors are kept. Moreover, the arousal axis is not actually a Big Five dimension: the paper says it uses Neuroticism as a motivating personality correlate, but operationalize the axis as affective arousal. Please justify these choices (e.g., with user studies or with the literature on linguistic markers of personality), and discuss why other common style dimensions, such as formality, verbosity, or informal writing (slang, typos), are not considered. This relates to the logical rigor of the paper. The current presentation of dimension choices seems more like an ad-hoc approach.
13 comments
AuthorsOpenReview ↗

Response to Reviewer b1hE

We sincerely thank you for your thorough and insightful review. We have carefully read and reflected on every comment, and we are grateful for your recognition of our research motivation, pilot study, and code release efforts.

We will do our best to address your concerns directly and head-on. If any concerns remain or we have misunderstood your comments, we would sincerely welcome further feedback.

AuthorsOpenReview ↗

R1 (W1/Q1). Fairness Motivation versus Robustness (1)

We agree that expression-style sensitivity is closely related to robustness, but our fairness perspective focuses on whether such sensitivity leads to systematic disparities in the quality of service users receive despite equivalent task requirements.

Q1.1: Is expression style treated as a protected attribute?

Our emphasis on examining fairness beyond traditional demographic attributes may have unintentionally given the impression that we regard expression style as another protected attribute. We apologize for this lack of clarity. We do not\textcolor{red}{\text{not}} treat expression style as a protected attribute, but rather as a linguistic characteristic of user instructions that serves as a grouping dimension for fairness evaluation. Specifically, PA-Style\texttt{PA-Style} operationalizes this characteristic into four controlled expression conditions, allowing us to examine systematic disparities in agent service quality across styles under equivalent task requirements.

More broadly, while fairness concerning protected attributes is a central paradigm, fairness is not exclusively defined in terms of such attributes. For example, individual fairness emphasizes that individuals who are similar with respect to a task should be treated similarly (Dwork et al., 2012). Our study draws on this broader principle to motivate fairness-relevant service-quality consistency across expression styles, rather than treating expression styles as protected demographic attributes.

Q1.2: Why study expression-style induced performance disparities as a fairness issue rather than solely a robustness issue?

We agree that expression-style induced performance variation falls within the broader scope of prompt sensitivity and robustness. However, our primary concern is not merely whether different expressions affect agent performance, but whether particular expression styles are associated with lower average task performance for a given agent under equivalent task requirements. General prompt sensitivity does not necessarily imply a consistent disadvantage for any particular style, whereas our study specifically investigates such style-related disparities across repeated evaluations.

The fairness implications arise when expression styles are considered in relation to users. Expression style is not merely an arbitrary input variation; it can also reflect how users habitually communicate their requests. If particular styles consistently lead to poorer agent performance, users who tend to express themselves in those styles may systematically receive lower-quality service despite having equivalent task requirements. We therefore examine these disparities from a fairness perspective, focusing on whether differences in user expression styles are associated with disparities in average agent task performance under equivalent task requirements.

Importantly, fairness and robustness are complementary rather than mutually exclusive perspectives. Improving overall task performance does not necessarily reduce cross-style disparities. Accordingly, we evaluate both Average Performance and Cross-style Disparity, aiming to improve service-quality consistency without sacrificing overall task effectiveness. Related research has similarly examined linguistic variation through both fairness and robustness perspectives (Lin et al., 2025).

Q1.3: Is there evidence that specific styles are strongly associated with specific user groups?

Prior research has documented associations between linguistic politeness and social factors (Danescu-Niculescu-Mizil et al., 2013; van Dorst et al., 2024). However, we do not claim that our four PA-Style\texttt{PA-Style} categories correspond to specific demographic or protected groups, nor that the observed disparities establish demographic discrimination. Such associations are not a prerequisite for investigating systematic service-quality disparities across expression styles, although further evidence would be necessary to support stronger claims about disparities affecting specific demographic groups.

We will clarify these distinctions and the scope of our fairness claims in the revised manuscript.

AuthorsOpenReview ↗

R1 (W1/Q1). Fairness Motivation versus Robustness (2)

W1: Novelty and Contributions Relative to Prior Work.

We sincerely appreciate your observation regarding prior studies on the effects of politeness and emotion on LLM performance. We acknowledge that our original novelty claims may have been interpreted too broadly, and we will clarify their scope to distinguish our agent-level contributions from previously established effects on LLM performance.

However, our contribution is not simply to demonstrate that politeness or emotion affects LLM performance, but to investigate whether and how expression-style sensitivity extends to the actual task execution of tool-using LLM agents, and how the resulting disparities can be mitigated.

WorkTargetStyleAgent ActionsMitigation Focus
Yin et al. (2024)LLMPolitenessNoNot studied
Dobariya & Kumar (2025)LLMPolitenessNoNot studied
Li et al. (2024)Generative AIEmotionNoPerformance enhancement
OursTool-using agentsPoliteness × ArousalYesCross-style disparity reduction (ISA)

① From LLM Responses to Agent Task Execution.

Prior studies primarily examine LLM response accuracy or generation quality, whereas tool-using agents additionally involve planning, decision-making, tool invocation, and environment interaction. Whether and how response-level style sensitivity extends to agent actions and end-to-end task completion cannot be directly inferred from these studies. We investigate this distinction by evaluating three tool-using agents and analyzing style-dependent tool-use behavior in Section 4.3.

② From Individual Style Dimensions to Controlled Compound-style Disparity Evaluation.

Rather than examining politeness or emotional cues individually, PA-Style\texttt{PA-Style} combines Politeness × Arousal into four controlled expression conditions. This enables systematic evaluation of cross-style performance disparities in agent task execution under equivalent task requirements, beyond isolated prompt-level performance effects.

③ From Performance-oriented Prompting to Disparity-oriented Mitigation.

Unlike prior approaches primarily aimed at improving model performance, ISA targets cross-style performance disparity reduction while preserving overall agent task performance. Since higher average performance does not necessarily imply smaller disparities, we evaluate both Average Performance and Cross-style Disparity to assess mitigation effectiveness.

We will revise the manuscript to better acknowledge prior work, clarify our contributions, and avoid overstating our novelty.

Reviewer_55FuOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates expression style-induced bias in LLM agents, examining how variations in user instruction styles affect task performance. The authors introduce PA-Style, a two-dimensional framework based on politeness and emotional arousal, to characterize different instruction styles. They further propose a semantic-preserving stylization framework to evaluate performance disparities and develop Instruction Style Alignment (ISA), a controllable text style transfer method designed to mitigate these disparities. Experiments on VitaBench demonstrate performance variations across instruction styles and evaluate the effectiveness of ISA.

Critical strengths

Interesting research perspective. The paper explores instruction expression style as a potential source of bias in LLM agents, extending fairness research beyond conventional demographic attributes.

Systematic evaluation framework. The proposed PA-Style framework provides a structured approach to examining different instruction styles, while semantic-preserving stylization attempts to control task-relevant information during evaluation.

Practical relevance. The study highlights the sensitivity of LLM agents to variations in user instructions and explores a lightweight style alignment approach to improve performance consistency across different expression styles.

Critical weaknesses

Weaknesses

  1. Insufficient theoretical justification for the psychological framework. The authors select only two dimensions from the Big Five personality model to construct PA-Style, but the rationale for this selection is not sufficiently convincing. In particular, the exclusion of the other three dimensions and the mapping from personality traits to expression styles lack adequate theoretical and empirical justification.

  2. Insufficient evidence of substantial style-induced performance disparities. The results in Table 1 do not convincingly demonstrate that stylized instructions lead to substantial or unacceptable performance degradation. LLMs are inherently sensitive to instruction variations, and different phrasings can naturally produce different responses. The observed discrepancies may therefore reflect general prompt sensitivity rather than a distinct expression style-induced fairness issue. More rigorous experiments are needed to isolate the effects of expression style from other instruction variations.

  3. Limited generalizability of the proposed mitigation method. Although Table 3 evaluates three LLM agents, Table 4 validates ISA only on Laguna-XS.2. Given that different agents exhibit different sensitivities to expression styles, evaluating a single agent is insufficient to establish the generalizability of ISA or rule out model-specific effects.

  4. Limited methodological novelty and insufficient comparisons. The proposed ISA is essentially a controllable text style transfer method built upon existing techniques. However, the experiments primarily compare ISA with prompt-based approaches, without including established text style transfer methods. Comparisons with representative style transfer and instruction normalization methods are necessary to demonstrate the novelty and practical necessity of ISA.

Questions for authors

No

6 comments
AuthorsOpenReview ↗

Response to Reviewer 55Fu

We sincerely thank you for your thoughtful and constructive review. We deeply appreciate the time and effort you devoted to examining our research motivation, evaluation framework, and mitigation method.

In the following response, we will do our best to address your concerns directly and head-on. If any concerns remain or we have misunderstood your comments, we would sincerely welcome further feedback.

AuthorsOpenReview ↗

RW1. Theoretical Justification and Dimension Selection of

① Why Big Five, and why select only two factors?

Our research motivation begins with user heterogeneity rather than textual variation alone. We aim to investigate whether users expressing equivalent task requirements in different styles may systematically receive different levels of service quality from LLM agents.

Prior research has established associations between personality traits and linguistic behavior (Mairesse et al., 2007; Ireland & Mehl, 2014). We therefore adopted Big Five as a structured, personality-informed starting point for identifying and selecting user expression characteristics, rather than choosing linguistic dimensions solely based on intuition.

Our selection of two factors was guided by two considerations: (1) Task-information preservation: whether the associated expression characteristics can be manipulated without changing task objectives, user requirements, or action-relevant constraints; and (2) Linguistic operationalizability: whether these characteristics can be explicitly, observably, and relatively independently operationalized in short, task-oriented user instructions.

Based on these considerations, we prioritized Interpersonal Stance, associated with Agreeableness, and Emotional Reactivity, associated with Neuroticism. The former characterizes how users formulate requests to an agent, while the latter concerns the emotional intensity conveyed in their expressions. These characteristics can be further operationalized as Politeness and Arousal, respectively, and combined into four controlled expression conditions.

In contrast, Openness-related expressions may involve user preferences, task choices, or novelty-seeking tendencies; Conscientiousness-related expressions may affect instruction details, completeness, or organization; and Extraversion-related characteristics, such as sociability and talkativeness, may be more difficult to isolate in short instructions without also varying verbosity or interaction patterns.

Thus, our selection does not imply that the remaining three factors are unimportant. Rather, we prioritized expression characteristics that can be manipulated within our controlled agent-task setting while preserving task-relevant requirements.

② Why operationalize Agreeableness as Politeness and Neuroticism as Arousal?

Our mapping does not directly equate personality traits with linguistic styles. Instead, it combines personality psychology with linguistic and affective theories to establish operational connections.

For Politeness, DeYoung et al. (2007) identify Politeness as an important aspect of Agreeableness, involving interpersonal tendencies such as respectfulness and avoidance of confrontation. These characteristics can be reflected in linguistic politeness strategies, including polite requests, indirect expressions, hedging, and imperative constructions (Danescu-Niculescu-Mizil et al., 2013).

For Arousal, emotional reactivity associated with Neuroticism, particularly the Volatility aspect discussed by DeYoung et al. (2007), provides a personality-related motivation for considering emotional intensity. Building on Russell's (1980) theory of affective arousal, we operationalize this characteristic along the Calm–Emotional dimension, using linguistic indicators such as emotional vocabulary, intensifiers, and exclamations (Argaman, 2010).

Therefore, Big Five provides a theoretical motivation for dimension selection ( Agreeableness and Neuroticism) from the perspective of individual differences, while linguistic and affective theories guide their concrete operationalization (Politeness and Arousal) in user instructions.


In summary, PA-Style\texttt{PA-Style} is a theoretically motivated framework that draws on interpersonal and emotional expression characteristics associated with Big Five and operationalizes them into controllable linguistic dimensions through linguistic and affective theories, rather than directly measuring personality traits.

AuthorsOpenReview ↗

RW2. Magnitude of Style-Induced Performance Disparities and Distinction from General Prompt Sensitivity (1)

We understand that your criticism involves two related questions: (1) whether the experimental results provide convincing evidence of substantial performance disparities, and (2) whether the observed disparities are specifically associated with expression styles rather than general prompt sensitivity. We answer both questions below.

① Magnitude of Observed Performance Disparities.

We would like to clarify that Table 1 reports a preliminary pilot study using limited, additive stylization, rather than a systematic manipulation of the full expression-style space. Specifically, we kept the original instruction's main body unchanged and introduced only a few modifications: prefixing it with "Hey there!", inserting several exclamation marks, and appending "Get this arranged right now! Don't dawdle!" The purpose was to preliminarily examine whether changes in instruction expression could affect agent task execution while preserving the original task requirements, rather than to characterize the full magnitude of performance disparities across different expression styles.

Under this limited-scope manipulation, Avg@4 decreased by 2.5 and 8.0 percentage points, Pass^4 decreased by 10.0 and 12.0 percentage points, and Pass@4 decreased by 4.0 and 6.0 percentage points for MiniMax-M2.7 and Qwen3.6-35B, respectively.

These pilot results provide an initial indication that even limited changes in instruction expression can affect agent task execution. However, they were not intended to quantify the full extent of performance disparities across different expression styles.

We subsequently introduced PA-Style to systematically examine four predefined compound expression styles. As shown in Table 3, MiniMax-M2.7 exhibits cross-style performance ranges of 10% Avg@4 and 18% on Pass^4.

Thus, Table 1 serves as an initial investigation of the existence of the problem, whereas Table 3 provides the systematic evaluation of cross-style performance disparities.

② Distinguishing Expression-style Disparities from General Prompt Sensitivity.

We agree that expression-style sensitivity falls within the broader scope of prompt sensitivity. We do not claim that expression-style sensitivity is a phenomenon entirely separate from general prompt sensitivity. Rather, our focus is on a structured and user-relevant form of input variation: whether different expression styles, under equivalent task requirements, are associated with differences in average agent task performance across the evaluated tasks.

Specifically, PA-Style defines four controlled conditions along Politeness × Arousal, allowing us to examine whether particular compound styles are associated with lower aggregate performance. Our concern is therefore not simply that different phrasings may produce different outcomes, but whether users who habitually communicate in certain expression styles may systematically receive lower average levels of agent service quality despite expressing equivalent task requirements.

③ Additional Experiment: Style-controlled Variation versus General Paraphrasing.

To examine whether the observed performance disparities exceed those induced by general phrasing variations, we conducted an additional controlled paraphrasing experiment using 50 tasks. For each task, we constructed two sets of four instructions from the same shared source instruction:

  • General Paraphrasing: Four paraphrases designed to preserve the original task requirements while introducing lexical and syntactic variations, without intentionally manipulating Politeness or Arousal. These paraphrases were not selected or filtered according to their PA characteristics.
  • PA-Style Variation: Four instructions corresponding to the predefined Politeness × Arousal conditions, also designed to preserve the original task requirements.

Both conditions used the same task set, agent configurations, number of input variants, four execution trials per variant, and evaluation protocol. Each agent completed all 1,600 planned executions, yielding 4,800 executions across the three agents.

We report Avg@4, the average success rate across four execution trials, and Pass^4, the proportion of tasks for which all four trials succeed. For each metric, we first aggregate each variant’s performance across tasks and then calculate the maximum-minus-minimum range across the four variants within each condition. The Range Difference is defined as PA-Style Range − Paraphrasing Range. All values below are expressed in percentage points.

Reviewer_xNMsOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigate the discrepancy of LLM performance due to differences in prompt politeness and emotional arousal. They find that different models prefer varying style combinations. Furthermore, the paper proposes a method to mitigate this expression style bias by training lightweight models to harmonize expression styles while preserving task information. Experiments show improvement from prompt engineering baselines.

Critical strengths

In addition to identify expression style bias, this work proposes an effective method to mitigate the bias. Their method addresses the problem by finding the superior expression style for an LLM and translate prompts into that style while making effort to preserve task structure and information. This seems a clever approach and their results show the effectiveness of the proposed method.

Critical weaknesses

The authors claim that this is the first paper "to identify user expression style as a novel bias dimension" in LLMs. However, as I learn about this field, this is not a very recent topic, especially on how politeness influence LLM performance. Prior papers use different terminology, but study similar questions. This paper therefore needs a more comprehensive review of prior work and a more precise justification of its contribution. Some closely related literature:

Yin, Z., Wang, H., Horio, K., Kawahara, D., & Sekine, S. (2024, November). Should we respect LLMs? A cross-lingual study on the influence of prompt politeness on LLM performance. In Proceedings of the Second Workshop on Social Influence in Conversations (SICon 2024) (pp. 9-35).

Dobariya, O., & Kumar, A. (2025). Mind your tone: Investigating how prompt politeness affects llm accuracy (short paper). arXiv preprint arXiv:2510.04950.

Li, C., Wang, J., Zhang, Y., Zhu, K., Wang, X., Hou, W., Lian, J., Luo, F., Yang, Q. & Xie, X.. (2024). The Good, The Bad, and Why: Unveiling Emotions in Generative AI. In Proceedings of the 41st International Conference on Machine Learning 235:28905-28934.

Questions for authors

In Section 3.2, I don't think it's fully motivated why the authors introduce the Big Five as the theoretically grounded framework. They also acknowledge that three of the five dimensions of this personality model doesn't apply to user expression style. Moreover, they don't operationalize the two remaining personality concepts but merely related expression styles. This makes the connection to personality model unnecessary, other frameworks in linguistic styles may address the problem more directly.

5 comments
AuthorsOpenReview ↗

Response to Reviewer xNMs

We sincerely thank you for your constructive feedback and the time and effort you have devoted to reviewing our work. We are particularly grateful for your recognition and encouragement regarding our mitigation approach. Your comments have prompted us to reconsider the novelty positioning and theoretical motivation of our paper more carefully.

We will do our best to address your concerns directly and head-on. If any of our explanations remain unclear or if we have misunderstood your comments, we would sincerely appreciate your further guidance.

AuthorsOpenReview ↗

R1-Weakness. Novelty and Relationship to Prior Work

Thank you for pointing out these closely related studies. We fully agree that prior work has already investigated the effects of Politeness and Emotion on LLM performance. We acknowledge that our original novelty claim may have been interpreted too broadly, and we will clarify its scope to distinguish our agent-level evaluation and mitigation contributions from previously established style effects on LLM performance.

We would like to clarify that our contribution is not\textcolor{red}{\text{not}} simply to discover that Politeness or Emotion affects LLM outputs, but to systematically \textcolor{red}{\text{systematically }}investigate whether and how these effects extend to end-to-end task execution by tool-using LLM agents, and to further evaluate and mitigate performance disparities across user expression styles.

WorkTargetStyleAgent ActionMitigation Focus
Yin et al. (2024)LLMPolitenessNoNone
Dobariya & Kumar (2025)LLMPolitenessNoNone
Li et al. (2024)Generative AIEmotionNoPerformance improvement
OursTool-using agentPoliteness × ArousalYesCross-style disparity mitigation

① From LLM Responses to Agent Task Execution.

We believe this is not a straightforward extension whose conclusions can be directly inferred from existing studies. Prior research has primarily focused on LLM response accuracy or text generation performance, whereas agents must additionally perform planning, decision-making, tool calls, and interactions with external environments. Therefore, whether style sensitivity observed at the response level persists at the action level, how it affects end-to-end task completion, and whether different agents exhibit similar style preferences remain questions that existing studies cannot directly answer.

For example, two users may express identical hotel-booking requirements using different expression styles. For a conventional LLM, style sensitivity may manifest as differences in generated responses. For an agent, however, it may further influence search strategies, tool-call sequences, and ultimately whether the booking is successfully completed, directly affecting the service users receive. Our work specifically investigates this issue by evaluating three tool-using agents and analyzing the effects of expression styles on tool-calling behavior in Section 4.3.

② From Individual Prompt Effects to Controlled Compound-style Disparity Evaluation.

Furthermore, our work goes beyond examining whether individual expression variations affect model performance. Instead, we systematically evaluate agent performance disparities across different expression styles from a user-centered\textcolor{red}{\text{user-centered}} perspective. Specifically, PA-Style\texttt{PA-Style} constructs four compound styles through Politeness × Arousal, enabling systematic comparisons of agent task performance under fixed task requirements and action-relevant constraints. This allows us to investigate whether systematic performance disparities are associated with particular expression styles, rather than merely observing performance fluctuations caused by individual prompt variations.

③ From Performance-oriented Prompting to Disparity-oriented Mitigation.

We recognize the value of prior research that improves model performance by modifying input expressions. For example, Li et al. (2024) propose EmotionPrompt, which primarily aims to improve generative AI performance through emotional stimuli.

In contrast, our mitigation objective is not limited to improving overall task performance. We also aim to reduce performance disparities across expression styles so that users with different expression styles may receive more consistent service quality from agents. To this end, we propose ISA\texttt{ISA}, which transforms user instructions into a relatively favorable, unified expression style based on agents' responses to different styles, while attempting to preserve task information and mitigate cross-style disparities.


In the revised manuscript, we will:

  • expand the Related Work to discuss relevant studies, particularly the three papers you highlighted;
  • narrow our novelty claims and revise the Abstract, Introduction, and Contributions accordingly.
AuthorsOpenReview ↗

R2-Question. Why Adopt Big Five Rather Than Existing Linguistic Style Frameworks? (1)

Thank you for this insightful question. Our original presentation may not have fully conveyed the distinct role of Big Five in motivating our dimension selection, potentially giving the impression that its introduction was unnecessary.

① Why Big Five rather than existing linguistic style frameworks?

We fully acknowledge that established linguistic frameworks, including Register Theory, Systemic Functional Linguistics (SFL), and Biber's Multidimensional Analysis (MDA), can characterize linguistic variation and expression styles.

However, our research motivation begins with user heterogeneity\textcolor{red}{\text{user heterogeneity}} rather than textual variation alone. Our goal is not\textcolor{red}{\text{not}} to exhaustively examine the many forms of input variation that may affect agent performance, but to investigate whether particular expression styles are associated with lower average task performance for a given agent under equivalent task requirements.

Therefore, beyond describing how language varies, we seek to ground our selection of expression-style dimensions in individual differences among users, rather than selecting linguistic features solely through intuition.

Prior research has established associations between personality traits and linguistic behavior (Mairesse et al., 2007; Ireland & Mehl, 2014). Since Big Five provides a well-established and structured framework for characterizing individual differences in personality, we adopted it as a theoretical starting point for identifying and organizing user expression characteristics.

To further clarify our rationale, we compare Big Five with several representative linguistic frameworks:

ComparisonRegister TheorySFLMDABig Five
Primary focusContext-dependent language variationLanguage functions and their relationships with contextDimensions of linguistic variation in corporaPersonality and individual differences
Main perspectiveContext-centeredFunction-centeredCorpus-drivenPersonality-informed
Support for dimension selectionDimensions may be selected based on communicative contextsDimensions may be selected based on linguistic functionsDimensions may be identified through corpus analysisMotivates dimension selection from the perspective of individual differences
Role and limitations in our studySupports style characterization but requires further selection of target dimensionsSupports linguistic operationalization but requires an explicitly defined style spaceOffers a data-driven alternative, depending on corpus availability and dimension interpretationMotivates dimension selection but does not directly define linguistic features

As illustrated above, these frameworks offer different but complementary perspectives on expression-style modeling. Register Theory emphasizes communicative contexts, SFL focuses on linguistic functions, and MDA identifies dimensions of linguistic variation through corpus analysis. Big Five, in contrast, provides a structured, personality-informed perspective on individual differences among users.

We recognize that linguistic frameworks provide more direct tools for characterizing and operationalizing linguistic expression styles. However, these frameworks do not\textcolor{red}{\text{not}}, by themselves, prescribe a set of expression-style dimensions grounded in individual differences among users. We therefore adopted Big Five not to replace these frameworks, but to draw on established relationships between personality and language use to provide a structured theoretical motivation for our user-centered dimension selection. We view these theoretical perspectives as complementary rather than mutually exclusive.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary

We sincerely thank all reviewers for their thoughtful and constructive feedback. We have carefully considered the concerns raised during the discussion and revised our manuscript to clarify the following points.

1. Fairness Motivation and Scope (b1hE, 55Fu).

Our fairness-oriented perspective concerns potential systematic disadvantages experienced by users with different habitual expression styles. From a user's perspective, if two users with different expression habits receive different levels of service despite having identical task requirements, this raises a fairness concern regarding unequal service quality. From an agent developer's perspective, the same phenomenon indicates that the agent is sensitive to variations in expression styles and lacks sufficient robustness. We therefore view fairness and robustness as complementary rather than mutually exclusive perspectives on the same phenomenon. Although our controlled experiments do not directly establish performance disparities across real-world user populations, they reveal expression-style-induced performance differences with potential fairness implications.

2. Novelty and Contributions (xNMs, 55Fu, b1hE).

We acknowledge prior research on the effects of politeness, emotional expression, and prompt sensitivity on model performance. We do not claim to be the first to identify the effects of linguistic variations on model behavior. Rather, we investigate this issue from the perspective of user expression styles and agent performance disparities. Specifically, we develop a controlled evaluation framework based on two dimensions, Politeness and Arousal, systematically examine how different expression styles affect agents' actual task execution, and further propose an agent-level mitigation approach. Thus, our work goes beyond examining whether linguistic variations affect model outputs to investigate how such effects manifest in agent task execution and how the resulting performance disparities can be mitigated.

3. PA-Style Design (xNMs, 55Fu, b1hE).

The Big Five provides a personality-informed starting point for identifying user-related expression differences, while linguistic and affective theories guide their operationalization. Specifically, the Politeness dimension is motivated by the interpersonal tendencies associated with Agreeableness and grounded in politeness theory, while the Arousal dimension draws on the emotional-reactivity aspect of Neuroticism and affective arousal theory. PA-Style operationalizes these two dimensions into controllable linguistic variations, without claiming to exhaust the full spectrum of human expression styles.

Manuscript Revisions. Based on the reviewers' constructive suggestions, we have made the following revisions:

  • Revised Abstract, Introduction, and Contributions: refined fairness framing and novelty claims;
  • Expanded Section 2: strengthened related work and comparisons with prior studies;
  • Revised Sections 3.2–3.3 and Table 2: clarified PA-Style's theoretical foundations and dimension selection;
  • Expanded Appendices A and C: clarified fairness motivation, linguistic foundations, and scope;
  • Updated Section 4.2 and Table 3: added repeated-trial variability estimates;
  • Added Appendix D.1 and Table 4: compared PA-Style variations with general paraphrasing;
  • Added Appendix D.2: evaluated consistency across independent evaluators;
  • Expanded Appendices E.1–E.2: added validation-failure analysis, cross-model validation, and human verification;
  • Updated Section 5.3, Table 5, and Appendix F.1: corrected ISA results and added fusion-weight sensitivity analysis.

We sincerely appreciate the reviewers' valuable feedback, which has helped us refine our research framing and clarify our contributions. We hope these clarifications and revisions will help address the reviewers' concerns and that they will consider the revised manuscript, its motivation, contributions, and findings with an open mind.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.