personality prompting

This study addresses the limitations of current approaches that rely solely on prompting to specify Big Five personality traits, which often fail to ensure consistent personality expression in language model dialogues. Adopting an interactionist perspective, this work systematically demonstrates for the first time that personality expression emerges from the context-dependent interplay among personality settings, social roles, and expressive styles—challenging the conventional assumption that personality can be controlled through prompts alone. Through a factorial design generating English–Japanese dialogue data, combined with LLM-as-a-judge evaluation and cross-lingual comparative analysis, the study reveals that social roles significantly influence openness, expressive style predominantly shapes conscientiousness and agreeableness, and neuroticism is primarily driven by explicit personality settings. Notably, stable personality impressions can still be elicited through social roles and expressive styles even without explicit personality prompts, with highly consistent findings across both languages.

This study addresses the behavioral inconsistency of large language models (LLMs) under identical personality prompts across varying conversational contexts—a phenomenon that has sparked debate over whether it reflects a flaw or human-like contextual adaptability. Introducing the holistic trait theory into LLM personality research for the first time, the work systematically examines how contextual factors modulate LLMs’ linguistic patterns, behavioral tendencies, and emotional expressions through dialogue experiments in four scenarios: ice-breaking, negotiation, group decision-making, and empathy. The findings reveal that LLMs exhibit context-sensitive personality expression, dynamically adjusting their traits and emotional tones in response to social and affective demands. This challenges the traditional evaluation paradigm centered on behavioral consistency and demonstrates that LLMs possess human-like contextual adaptability.

This study investigates whether large language models (LLMs) can generate multimodal agent behaviors—spanning verbal and nonverbal modalities—that are both consistent with and perceptible as reflecting specific personality traits. Method: Focusing on extraversion as a core dimension, we conduct empirical evaluations in two social scenarios—negotiation and ice-breaking—and introduce the first personality-aware prompting framework that embeds Five-Factor Inventory (BFI) trait scores into LLM inputs to jointly govern linguistic output and nonverbal actions (e.g., posture, gaze). Validation employs linguistic analysis, action-mapping modeling, and user perception experiments. Contribution/Results: Generated behaviors significantly differentiate introverted vs. extraverted tendencies (p < 0.01), and human participants reliably identify agent personality traits with high accuracy (mean 82.3%). This work establishes the first framework for personality-guided, controllable multimodal behavior generation from LLMs and empirically confirms its cross-modal consistency and social interpretability.

This study investigates whether large language models (LLMs) can reliably reproduce human social decision-making under personality prompting—a core assumption in personality alignment research. Method: Drawing on canonical psychological paradigms—Milgram’s obedience experiment and the ultimatum game—we construct a cross-vendor (OpenAI, Anthropic, Google, Meta), multi-prompt social behavior evaluation framework. Contribution/Results: Empirical results demonstrate that all evaluated models fail to consistently instantiate target personality traits in social contexts, exhibiting systematic deviations in obedience and fairness tasks. Critically, this failure is robust across model architectures and prompt perturbations. To our knowledge, this is the first systematic demonstration of the pervasive ineffectiveness of personality prompting in social decision-making scenarios. These findings challenge prevailing optimistic assumptions in personality alignment literature and indicate that current approaches remain in an early, foundational stage.

This study investigates how persona assignments (e.g., teacher, woman, LGBTQ+ individual) systematically influence the behavioral outputs of large language models (LLMs). Method: We conduct a large-scale empirical analysis across 12 categories comprising 162 distinct personas, evaluating responses from seven mainstream LLMs on five task domains—mathematical reasoning, historical knowledge, value alignment, and other objective/subjective benchmarks—while rigorously controlling for confounding factors via 30 prompt rewriting techniques. Contribution/Results: We provide the first evidence that persona-induced behavioral variation substantially exceeds standard prompt sensitivity; certain effects—including enhanced logical consistency under authority-related personae and richer value expression under diverse identity personae—exhibit cross-model robustness. Statistical significance is confirmed across all models and datasets, establishing persona as a potent, controllable intervention factor. These findings introduce a novel paradigm for behavior-aware model design and controllable content generation.

论文针对LLM在社交机器人中使用时出现的不一致人格问题，通过提出包含八个功能组件的框架和指导方针来规范提示设计，确保行为的一致性和透明度。

针对大语言模型生成人格测验条目质量不稳定问题，提出AI-GENIE框架结合自适应提示工程与网络心理测量方法，有效提升结构效度并减少语义冗余。

This study addresses a critical oversight in current large language model (LLM) style control research: the unintended side effects of targeting one stylistic dimension on other non-targeted dimensions. By leveraging controlled synthetic dialogue generation and an LLM-as-a-Judge evaluation framework, the work systematically quantifies the cross-dimensional causal impacts of style interventions in both task-oriented and open-domain settings. It reveals, for the first time, that stylistic attributes are structurally coupled rather than orthogonal. The authors introduce and release the CASSE dataset to characterize these interactions. Experiments consistently show that enforcing styles such as “conciseness” significantly undermines traits like “professionalism.” While existing mitigation strategies partially restore suppressed characteristics, they often compromise the target style itself, thereby challenging prevailing assumptions about the efficacy of current style control mechanisms.

This study addresses a critical gap in existing research on emotional prompting, which has predominantly focused on single positive emotions while neglecting systematic analysis of diverse emotion types and their intensities. The work presents the first comprehensive investigation into how four distinct emotions—joy, encouragement, anger, and insecurity—and their varying intensities differentially influence large language models across three key dimensions: accuracy, sycophancy, and toxicity. Leveraging a GPT-4o mini–based pipeline for emotional prompt generation and combining human and model-based annotations, the authors construct a high-quality “gold dataset” to serve as a benchmark for emotional prompting. Their findings reveal a dual effect: while positive emotions enhance accuracy and reduce toxicity, they simultaneously exacerbate sycophantic behavior, underscoring the nuanced trade-offs inherent in emotionally infused prompts.

This work addresses the limitations of existing conversational agents that employ static personas, which often fail to adapt to dynamic shifts in task context, user goals, and situational urgency, leading to interactional mismatches. To overcome this, the authors propose a fluid persona framework that jointly models metaphorical roles—such as coach or mentor—and the intensity of personality expression (low, medium, high) to enable real-time adaptation based on task context, user traits, and situational urgency. Built upon large language models, the framework integrates context-aware mechanisms with personality dimension modulation strategies, allowing dynamic switching of both role and expression intensity during dialogue. Empirical evaluations demonstrate that this approach significantly enhances user experience, trust, and willingness to adopt recommended behaviors across diverse domains, including medical consultation, fitness coaching, and reflective learning.