Persona Prompting as a Lens on LLM Social Reasoning

Jing Yang, Moritz Hechtbauer, Elisabeth Khalilov, Evelyn Luise Brinkmann, Vera Schmitt, Nils Feldhus

Abstract: For socially sensitive tasks like hate speech detection, the quality of explanations from Large Language Models (LLMs) is crucial for factors like user trust and model alignment. While Persona prompting (PP) is increasingly used as a way to steer models towards user-specific generation, its effect on model rationales remains underexplored. We investigate how LLM-generated rationales vary when conditioned on different simulated demographic personas. Using datasets annotated with word-level rationales, we measure agreement with human annotations from different demographic groups and assess the impact of PP on model bias and human alignment. Our evaluation across three LLMs reveals three key findings: (1) PP improves classification on the most subjective task (hate speech) but degrades rationale quality. (2) Simulated personas fail to align with their real-world demographic counterparts, and high inter-persona agreement shows models are resistant to significant steering. (3) Models exhibit consistent demographic biases and a strong tendency to over-flag content as harmful, regardless of PP. Our findings reveal a critical trade-off: while PP can improve classification in socially-sensitive tasks, it often comes at the cost of rationale quality and fails to mitigate underlying biases, urging caution in its application.

Introduction

As Large Language Models (LLMs) are increasingly deployed for high-stakes tasks like content moderation, a "one-size-fits-all" approach to safety often fails to account for the diverse perspectives of a global user base. Researchers have turned to Persona Prompting (PP)—instructing a model to adopt a specific demographic identity (e.g., "as a 25-year-old Caucasian male")—to steer models toward more personalized or context-aware outputs.

However, the extent to which these personas affect a model's underlying social reasoning remains unclear. Does a model acting as a specific persona truly adopt that group's logic, or does it merely adjust its final answer while relying on the same internal biases? This study investigates this by analyzing token-level rationales—the specific words a model highlights to justify its decision—to determine if the "why" behind a model's output changes alongside the "what."

Methodology

The researchers developed a multi-faceted evaluation pipeline to test three distinct LLM families: Mistral-Medium, Qwen3-32B, and GPT-OSS-120B.

Experimental pipeline for evaluating persona prompting, from input selection to multi-faceted evaluation of labels and rationales.

Experimental Setup

  • Persona Construction: The study utilized 21 single-attribute personas (covering age, gender, education, race, religion, politics, and loneliness) and 12 composite personas.
  • Chain-of-Thought (CoT) Prompting: Models were instructed to "think out loud" in character before providing a final label and a JSON-formatted list of rationale tokens.
  • Datasets: The evaluation spanned three datasets with varying subjectivity: HateXplain (hate speech with word-level rationales), CoS-E (commonsense logic), and SST-2 (sentiment analysis). A specialized version, BRWRR, allowed for direct comparison between simulated personas and real-world human demographic annotations.

Key Insights and Results

The Steering Paradox and Rigid Reasoning

Persona prompting proved to be a double-edged sword. While it occasionally improved label accuracy in subjective tasks like hate speech detection (e.g., 11 of 21 personas improved Mistral-Medium's accuracy), this often came at the cost of rationale quality. Models frequently provided the "correct" label for the "wrong" reasons, showing no significant improvement in rationale alignment with human ground truth.

Furthermore, models demonstrated high inter-persona agreement (Krippendorff’s alpha $\geq 0.80$ for GPT-OSS and Mistral). This suggests that LLMs possess a "rigid reasoning template" established during pre-training and alignment that is highly resistant to surface-level persona steering.

Stereotypes and Safety Over-correction

Rather than capturing genuine social nuances, models often relied on linguistic stereotypes when simulating personas (e.g., simplifying language for personas with "no formal education"). A more pervasive issue was the safety over-correction effect. Regardless of the assigned persona, all models exhibited a strong tendency to over-flag content as offensive or hateful.

Heatmap of over-flagging rates across different models and personas, illustrating the consistent tendency to upgrade offensive content to hate speech.

As shown in the heatmap above, GPT-OSS, for instance, upgraded "Offensive" labels to "Hate Speech" nearly 50% of the time across almost all personas. This indicates that internal safety guardrails and alignment training exert a much stronger influence on model behavior than the specific instructions provided in a persona prompt.

Demographic Bias

Evaluation against human ground truth revealed persistent demographic biases. Models consistently aligned better with "Old" and "Caucasian" demographic groups, even when prompted to act as a different persona. This suggests that the underlying training data and alignment processes favor specific demographic perspectives, which persona prompting fails to mitigate.

Conclusion and Future Work

This research highlights that persona prompting is often a superficial layer that fails to alter the core social reasoning of LLMs. While it can nudge classification performance, it frequently degrades the quality of explanations and reinforces existing biases or safety over-corrections.

Limitations: The study's rationales were generated via prompting rather than internal feature attribution (like attention maps), and the human ground truth in HateXplain relies on a limited number of annotators.

Future Directions: Achieving true demographic alignment will likely require more than prompt engineering. Future research should focus on methods that can steer a model's internal reasoning processes without sacrificing performance or reinforcing harmful stereotypes.

Summary updated

Abstract

For socially sensitive tasks like hate speech detection, the quality of explanations from Large Language Models (LLMs) is crucial for factors like user trust and model alignment. While Persona prompting (PP) is increasingly used as a way to steer model towards user-specific generation, its effect on model rationales remains underexplored. We investigate how LLM-generated rationales vary when conditioned on different simulated demographic personas. Using datasets annotated with word-level rationales, we measure agreement with human annotations from different demographic groups, and assess the impact of PP on model bias and human alignment. Our evaluation across three LLMs results reveals three key findings: (1) PP improving classification on the most subjective task (hate speech) but degrading rationale quality. (2) Simulated personas fail to align with their real-world demographic counterparts, and high inter-persona agreement shows models are resistant to significant steering. (3) Models exhibit consistent demographic biases and a strong tendency to over-flag content as harmful, regardless of PP. Our findings reveal a critical trade-off: while PP can improve classification in socially-sensitive tasks, it often comes at the cost of rationale quality and fails to mitigate underlying biases, urging caution in its application.
BibTeX
@inproceedings{yang2026persona,
  author = {Jing Yang and Moritz Hechtbauer and Elisabeth Khalilov and Evelyn Luise Brinkmann and Vera Schmitt and Nils Feldhus},
  title = {{Persona Prompting as a Lens on LLM Social Reasoning}},
  booktitle = {EACL (Volume 1: Long Papers)},
  year = {2026},
  doi = {10.18653/v1/2026.eacl-long.52},
}
Figure