Persona Prompting as a Lens on LLM Social Reasoning

TL;DR

This work critically examines Persona Prompting (PP) as a lens to study LLM social reasoning in sensitive tasks. Using HateXplain and BRWRR datasets across three LLMs, the authors quantify both label decisions and token-level rationales, highlighting a task- and model-dependent landscape where PP can improve subjectively difficult label prediction but often harms rationale quality and does not reliably reduce demographic biases. Key findings show high inter-persona agreement and persistent guardrails, with simulated personas frequently failing to match real-world demographics. The results argue for caution in deploying PP for alignment and call for approaches that steer underlying reasoning without sacrificing performance or reinforcing biases. The study contributes a rigorous auditing framework for PP and emphasizes the need for robust, outcome-aligned methods beyond persona-based customization.

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.

Figures (8)

• Figure 1: Our pipeline: Datasets are combined with persona prompts, fed to LLMs, and the resulting labels and rationales are evaluated against ground truth, inter-persona agreement, and human demographic groups.
• Figure 2: Label prediction MAE scores $\downarrow$ of baseline (no persona) and different single-attribute personas on HateXplain. Error bars represent 95% confidence intervals (CI) with bootstrapping resampling. If the line does not cross the baseline, the difference is significant.
• Figure 3: Over-flagging rate across different labels. N: Normal, O: Offensive, H: Hate speech.
• Figure 4: Rationale Token-$F_1$$\uparrow$ of baseline (no persona) and different single-attribute persona performances on HateXplain, excluding rationales from the "Normal" label. Error bars incidates 95% CIs, significantly different persona results are indicated with squares.
• Figure 5: Performance (left: accuracy/Macro-$F_1$; right: Token-$F_1$) for (no persona) baseline and personas across demographic groups. BY: African American Young, WY: Caucasian Young, LY: Hispanic Young, BO: African American Old, WO: White Old, LO: Hispanic Old. Each group has its own ground truth labels and rationales. Error bars incidates 95% CIs, significantly different personas are marked with a dark gray cycle around.