Personality, Role, and Expressive Style in Large Language Models: An Interactionist Analysis

Abstract: Prompt-based personality control is a key technique for designing LLM dialogue agents that behave consistently across social contexts. However, specifying Big Five personality traits (BFTs) in a prompt does not ensure that the intended traits are expressed in generated utterances. This paper investigates this mismatch from an interactionist perspective, viewing personality expression as a context-dependent outcome shaped by the interplay between trait specification and situational factors. We analyze how perceived BFT expression in LLM-generated dialogue is influenced by three prompt factors: personality traits, dialogue roles, and expressive styles. Using a factorial design that combines six personality conditions, three roles, and three expressive-style conditions, we generate 1,080 LLM-agent dialogues in each of English and Japanese. We then evaluate the target agent's utterances using an LLM-as-a-judge framework to estimate expressed Big Five traits. The results show that expressed personality is shaped not only by explicit trait specification, but also by dialogue role and expressive style. These effects are trait-specific: dialogue role strongly influences Openness, expressive style substantially shapes Conscientiousness and Agreeableness, and explicit trait specification dominates Neuroticism. Even without explicit personality-trait specification, social and expressive conditions induce distinct personality-like impressions. Cross-linguistic comparisons show broadly similar patterns between English and Japanese dialogues, with noticeable differences only under specific combinations of personality, role, and expressive style. These findings suggest that personality control in LLM agents should be understood not as a direct consequence of trait prompting, but as a context-dependent process involving personality specification, social role, and expressive style.

• The paper uses a factorial study of 1,080 GPT-5.2 dialogues per language to show that expressed personality depends jointly on Big Five prompts, roles, and expressive styles.
• The paper finds that role most strongly shapes Openness, expressive style dominates Conscientiousness and Agreeableness, and explicit prompting overwhelmingly controls Neuroticism, with effects varying by context.
• The paper shows broadly similar English–Japanese patterns but cautions that LLM-based ratings measure perceived expression rather than validated personality and require human-judgment replication.

Motivation and framework

Prompt-based personality control is a widely used technique for shaping LLM dialogue agents, yet prior work has shown that specifying Big Five traits (BFTs) in a prompt does not reliably produce the intended trait expression in generated utterances. This paper addresses that mismatch from an interactionist perspective, drawing on trait theory traditions in which behavior is an outcome of dispositions interacting with situational constraints rather than a direct readout of internal traits. The central claim is that perceived personality in LLM dialogue is jointly determined by three prompt factors—explicit BFT specification, the agent's assigned role, and its expressive style—and that these factors cannot be meaningfully analyzed in isolation.

The study operationalizes this claim with a full factorial design crossing six personality conditions (Unspecified plus one high-trait condition for each of Openness, Conscientiousness, Extraversion, Agreeableness, and Neuroticism), three roles (Chat, Salesperson, Customer in a consumer-electronics retail scenario), and three expressive styles (Unspecified, Emotional, Rational). Dialogues were generated with GPT-5.2 between dyads of a target agent and an interlocutor agent, 20 dialogues per condition per language, yielding 1,080 dialogues each in English and Japanese. Perceived BFTs were scored on five-point Likert scales by Gemini 2.5 Flash as judge, with a separate evaluator model chosen to mitigate self-preference bias.

Evaluation robustness

Before the main analyses, the authors assessed inter-evaluator consistency by re-scoring English dialogues with OpenAI o3-mini. Condition-level RMSE ranged from 0.275 (Neuroticism) to 0.568 (Extraversion), with individual-dialogue Pearson correlations of 0.504–0.903. This supports that the trait estimates are not strongly dependent on the particular LLM judge. The authors are explicit, however, that this establishes only consistency among LLM judges, not equivalence to human judgments; all scores should be read as estimates of perceived personality expression rather than validated psychometric measurements. This caveat conditions every downstream result.

Trait specification alone is insufficient

Under the Chat role baseline with Unspecified expressive style, specifying a trait generally raised the corresponding expressed score, most clearly for Openness and Neuroticism and less so for Conscientiousness and Agreeableness. The effect is therefore real but heterogeneous across dimensions—an early indication that personality prompting is not a uniform control mechanism.

More striking is what happens without any trait specification: role assignment alone shifted expressed profiles. Under the Unspecified personality condition, Conscientiousness was higher in Salesperson and Customer roles than in Chat, while Openness was higher in Chat than in the retail roles. Purchase-oriented interaction thus induced goal-directed, responsibility-oriented impressions, whereas unconstrained chat induced flexible, exploratory ones. Personality-like impressions emerge from social context even absent explicit trait prompts—a result with direct implications for agents whose prompts specify only roles or tasks.

Quantifying the joint effects

Two-way ANOVAs (Personality × Role; Personality × Expressive Style) and three-way ANOVAs over all factors reveal strongly trait-specific structure:

Trait	Dominant factor (three-way ANOVA, $\omega^2$)	Notable secondary effects
Openness	Personality (.305); Role (.247)	Personality × Role .141
Conscientiousness	Expressive Style (.408)	Role × Style .071
Extraversion	Expressive Style (.274); Personality (.212); Role (.191)	All three substantial
Agreeableness	Expressive Style (.402)	Personality .117
Neuroticism	Personality (.685)	All others ≤ .076

Several results stand out quantitatively. For Openness under the two-way analysis, Role ($\omega^2 = 0.424$) actually exceeded Personality ($\omega^2 = 0.254$)—the assigned role mattered more than the trait prompt itself. For Conscientiousness and Agreeableness, expressive style dominated ($\omega^2 = 0.638$ and $0.454$ respectively in two-way analyses): Rational style increased perceived Conscientiousness while decreasing Extraversion and Openness; Emotional style increased Agreeableness while decreasing Conscientiousness and Emotional Stability. For Neuroticism, explicit specification was overwhelmingly dominant ($\omega^2 = 0.926$ in the two-way analysis), producing strong, stable expression across contexts. Multidimensional scaling configurations corroborate these patterns exploratorily: under the Salesperson role, profiles for all personality conditions except Neuroticism cluster tightly, indicating role-induced convergence of personality expression.

The three-way interactions were statistically significant but modest relative to main effects and major two-way interactions. The paper's key structural finding is accordingly not that "everything interacts," but that the identity of the dominant factor differs across Big Five dimensions. Practically, this means a trait prompt effective in one setting may be amplified, suppressed, or redirected elsewhere—for example, Emotional style reduced expressed Conscientiousness under the Chat role but not under Salesperson/Customer roles, and enhanced expressed Openness specifically in the Customer role.

Cross-linguistic comparison

English–Japanese comparisons showed broadly similar patterns, with per-trait RMSE of 0.30–0.50 across all 54 condition combinations. Differences concentrated in specific cells: in the Unspecified–Chat–Emotional condition, emotionally expressive chat was judged more neurotic in English than Japanese; in the Neuroticism–Salesperson–Rational condition, emotional stability remained higher in Japanese, indicating attenuation of Neuroticism expression under that particular combination. Language thus modulates personality expression condition-dependently rather than uniformly. One methodological caveat applies here: the same English evaluation prompt was used to score both languages, which may bias judgments of Japanese utterances.

Relation to prior work

Prior studies have examined Big Five prompting, persona/role assignment, style transfer, and affective stimuli such as EmotionPrompt, but typically one factor at a time. The contribution here is the factorial combination of trait, role, and style with LLM-judged evaluation of expressed traits in dialogue, grounded explicitly in interactionist personality theory (Mischel; Fleeson's density-distribution and Whole Trait Theory accounts). The discussion also connects the findings to latent-space steering approaches: if expression depends on context, steering a single personality dimension in isolation may not yield stable cross-contextual behavior—an open empirical question the paper does not resolve.

Limitations

The paper concedes several boundaries plainly. Evaluation relies entirely on LLM judges without human annotation, so scores are estimates of perceived expression, not validated measurements; replication across more generation and evaluation models is needed. The design space is narrow—one retail domain for the constrained roles, two expressive styles, single prompt formulations per factor—so "role" effects may partly reflect task structure or domain norms. The cross-linguistic comparison covers only English and Japanese with a shared English evaluation prompt, and dialogues are agent–agent rather than human–LLM. Finally, ANOVA observations are repeated stochastic generations under identical prompts, not independent participants, so inferential statistics describe variation across generated samples.

Conclusion

This study demonstrates that expressed Big Five traits in LLM-generated dialogue are context-dependent outcomes of trait specification, role, and expressive style acting jointly, with the dominant factor differing systematically across trait dimensions: role shapes Openness, expressive style dominates Conscientiousness and Agreeableness, and explicit specification overwhelmingly determines Neuroticism. Personality-like impressions arise even without trait prompts, and cross-linguistic differences appear only under specific factor combinations. The practical implication is that personality-controlled agents must be designed and evaluated under the concrete social and expressive conditions of deployment, treating personality, role, and style as coupled control variables. The principal open question left by the paper is whether these joint contextual effects are reflected in latent representations and whether representation-level control can achieve context-robust personality expression.