Source-linked AI summary
Persona Prompting as a Lens on LLM Social Reasoning
Jing Yang, Moritz Hechtbauer, Elisabeth Khalilov, Evelyn Luise Brinkmann, Vera Schmitt, Nils Feldhus
TL;DR
Persona prompting is intended to make LLM decisions and explanations better reflect diverse social perspectives, but its effects on rationales remain underexplored. The paper evaluates persona-conditioned labels and word-level rationales across three tasks and three LLMs. It finds that prompting can improve classification on hate speech while worsening rationale selection and failing to align models with corresponding demographic groups. Persistent demographic biases and over-flagging remain, motivating caution about automated detection tools.
Problem
Persona prompting’s effect on LLM-generated rationales and alignment with diverse human perspectives remains underexplored in socially sensitive and subjective tasks.
Method
The paper evaluates demographic persona prompts across three datasets and three LLMs, comparing labels and token-level rationales with ground truths, human demographic groups, and inter-persona agreement.
Results
Persona prompting improves label prediction on HateXplain but produces no rationale-selection gains and mostly worsens rationale performance there, while effects on other tasks are slightly negative or negligible.
Takeaways & Limitations
Persona prompting can trade improved classification on socially sensitive tasks for poorer rationales and does not reliably mitigate demographic misalignment or bias.
Takeaways & Limitations
The study uses majority-vote ground truths from only three HateXplain annotators, prompted rationales that may not reflect internal activations, and three representative models whose behaviors may not generalize.
Abstract
from arXiv · showhide
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.
1 Introduction
This paper examines how demographic persona prompting affects LLM classifications and token-level rationales in socially sensitive and subjective tasks. It finds that persona prompting can improve hate-speech classification while worsening rationale quality and failing to remove persistent demographic biases.
- Motivation and approach: The study links persona-conditioned behavior to both classification outputs and token-level rationale selection.It evaluates agreement with ground truths and human annotations from different demographic groups.
- Motivation and approach: The evaluation uses three datasets with different levels of subjectivity to examine how demographic prompts influence labels and rationales.The authors frame this as an audit of model social reasoning and internal social representations.
- Main findings: Persona prompting has task-dependent effects: it can improve hate-speech label prediction but often harms rationale quality.The paper distinguishes classification performance from rationale alignment rather than treating them as equivalent.
- Main findings: LLMs retain consistent demographic biases and tend to over-flag normal content as hate speech despite persona prompting.The authors characterize this behavior as guardrail-aligned over-flagging.
2 Related Work
Related work presents persona prompting as a scalable way to tailor LLM outputs, while emphasizing risks such as bias, stereotyping, and reduced factual consistency. This paper extends that literature by studying fine-grained rationale alignment across socially sensitive and subjective tasks.
- Persona prompting: Persona prompting can generate tailored outputs for diverse viewpoints and has been applied to social science and pluralistic-alignment research.Prior studies also report measurable behavioral effects and improvements in subjective text quality.
- Risks and reliability: Persona simulation can amplify biases, reduce factual consistency, and introduce stereotypical reasoning, especially in sensitive domains such as toxicity detection.These risks motivate work on persona coherence and psychological grounding.
- Research gap: Most prior studies emphasize classification or high-level explanation quality rather than fine-grained rationale alignment.The present work addresses this gap in hate-speech detection and less overtly social subjective tasks.
3 Evaluation Design
The evaluation combines demographic persona prompts with three tasks and compares model labels and rationales against human annotations, ground truths, and inter-persona consistency. It uses multiple persona designs, three LLMs, task-specific metrics, repeated runs, and bootstrap confidence intervals.
- Persona design: The study uses single-attribute personas and composite age, gender, and ethnicity personas to simulate demographic viewpoints.The composite personas are designed for comparison with six real demographic annotation groups.
- Prompting procedure: Each role-based prompt asks a model to classify an input in character and return a label plus a word-level rationale in structured JSON.The prompts include a persona description, task and input text, step-by-step reasoning, and a rule to remain in character.
- Baselines: Neutral prompts omit persona context while retaining the classification, reasoning, and structured-output format as the baseline.The study also tests a prompt variant without step-by-step reasoning, which performs worse, particularly for rationale selection.
- Datasets and tasks: The experiments cover HateXplain hate-speech detection, CoS-E commonsense reasoning, and SST-2 sentiment analysis.These tasks span socially sensitive, subjective, and less subjective settings.
- Experimental procedure: Three LLMs are evaluated across three independent runs, with 95% confidence intervals estimated from 1,000 bootstrap iterations for persona-baseline differences.The model set includes Mistral-Medium2, Qwen3-32B, and GPT-OSS-120B.
- Evaluation metrics: Classification is evaluated with accuracy, Macro-F1, MAE, and ME, while rationale selection uses token-level Intersection-Over-Union.The metrics are assigned according to task properties, including ordinal labels for HateXplain.
- Consistency analysis: Krippendorff’s α measures inter-persona consistency for both labels and binary rationale arrays.For rationales, each token is treated as a binary annotation task.
4 Results
Persona prompting affects label prediction differently across models and tasks, while rationale alignment generally does not improve and inter-persona agreement remains high. Models also show demographic bias and systematically over-flag harmfulness, with political-view personas showing especially low agreement.
- Label Prediction Over-flagging: All models produced positive Mean Error values, indicating systematic over-flagging of statements as Offensive or Hate speech when the ground truth was less severe.The study visualizes these transitions as over-flagging rates across Normal, Offensive, and Hate speech labels.
- Label Prediction Over-flagging: GPT-OSS had the highest O → H over-flagging rates, Mistral was high on N → O and O → H, and Qwen3 was lowest on N → H and O → H but high on N → O.The authors report that top-performing personas, including Male, Not lonely, Right-wing, White, and Atheist, had lower over-flagging rates.
- Label Agreement Across Personas: Inter-persona label agreement was high for GPT-OSS and satisfactory for Mistral, but Qwen3 remained below moderate agreement across all categories.Political view was the lowest-agreement category for GPT-OSS and Mistral, while Qwen3 also showed low agreement for education and loneliness.
- Persona Agreement on Rationales: Rationale agreement was slightly lower than label agreement for GPT-OSS and Mistral but high overall, while Qwen3 approached the tentative agreement threshold.Across 12 personas, agreement was very high for all models, with Qwen3 showing lower agreement specifically for rationale extraction.
- Qualitative analysis of Chain-of-Thought reasoning: Persona-conditioned reasoning produced distinct linguistic styles and analytical frameworks that led to different classifications for the same input.Higher education recognized an attacked protected group as Hate speech, while No formal education and Left-wing produced Offensive language and Right-wing produced Normal.
- Rationale Selection: Persona prompting failed to align rationale selection with demographic annotator groups, and models showed much higher alignment with the Caucasian annotator group.Models also performed better overall on Caucasian groups, indicating persistent demographic bias with and without personas.
- Label Prediction: 11 of 21 Mistral-Medium personas significantly improved HateXplain label prediction, whereas 16 of 21 Qwen3-32B personas significantly degraded it.GPT-OSS-120B showed the least deviation from its baseline, demonstrating model-dependent susceptibility to persona steering.
- Rationale Selection: Persona prompting did not improve HateXplain rationale selection despite label-accuracy gains, and Mistral and Qwen3 personas generally scored below their baselines on rationale Token-F1.GPT-OSS personas remained within the baseline performance range, while White and Asian race personas were the only noted exceptions for Mistral and Qwen3.
5 Discussion
Persona prompting has task-dependent effects: it can improve label prediction on HateXplain while failing to improve, and often worsening, rationale selection. Simulated personas also show weak demographic alignment, while persistent biases and over-flagging remain across prompting conditions.
- PP improves label prediction on HateXplain but has slightly negative effects on CoS-E and SST-2.
- PP produces no rationale-selection gains and mostly worsens performance on HateXplain.
- HateXplain’s greater subjectivity and social sensitivity produce more classification variability under PP than the less ambiguous CoS-E and SST-2 tasks.
- High inter-persona agreement and similar top-performing personas indicate weak alignment between simulated personas and real-world demographic groups.
- Across datasets, demographic bias and guardrail-aligned over-flagging persist regardless of PP.
6 Conclusion
The paper studies whether socio-demographic persona prompts reveal and steer LLM social reasoning through word-level rationales. Across three LLMs and tasks, PP improves label performance on hate speech but worsens rationale selection, fails to align models with demographic groups, and leaves bias and over-flagging intact.
- PP improves label performance on the more subjective hate-speech task but worsens rationale selection on that task.
- PP usually fails to steer models toward better alignment with corresponding demographic groups.
- Regardless of personas, models perform better on White, African American, and Old groups and heavily over-flag harmful content.
Limitations
The study’s conclusions are bounded by limitations in its ground truths, rationale-generation approach, and model selection.
- HateXplain uses three annotators per sample with majority-vote resolution, which may miss the full spectrum of interpretations.
- Prompt-generated rationales may not faithfully reflect models’ internal activations because they are not produced through traditional feature attribution.
- Only three representative models were evaluated, and Qwen3’s divergence suggests other LLMs may behave differently.
Ethical Statement
The study aims to support safer and fairer models while warning that incorrect, variable outputs can harm reliability and user judgment.
- Incorrect classifications can reduce system reliability and potentially affect users negatively.
- Outputs may vary substantially with the simulated persona or language model, so excessive trust in automated detection tools may reduce careful user judgment.
Contributions
The project was jointly conceptualized and supervised, with implementation, experimentation, analysis, and writing distributed across the author team.
- JY and NF conceptualized the project and jointly supervised the work.
- MH, EK, and ELB provided the initial implementation and writing for the course project.
- JY and MH conducted further experiments, while JY and NF performed further analyses.
A Experimental details
The experiments construct demographic personas across several attributes and evaluate LLM classifications and rationales under structured prompting, with and without chain-of-thought instructions.
- Persona construction: The study constructs personas using age, gender, race, education, political view, religion, and loneliness.The attributes are motivated by demographic grouping, intellectual exposure, controversy, worldview, and social sensitivity considerations.
- Experimental setup: Each persona receives one prompt per instance, producing 31,500 HateXplain, 18,000 CoS-E, and 9,468 SST-2 completions per model, besides baseline generations.The counts use 500 HateXplain instances, 500 CoS-E instances, and 263 SST-2 instances, each with three runs.
- Prompting and outputs: Persona prompts place persona-related content in the user prompt because the authors found models acted more like the persona there than in the system prompt.
- Prompting and outputs: Prompts ask models to predict labels, select supporting rationale words, and provide reasoning in a constrained JSON format.HateXplain uses three labels and word-level rationale selection; CoS-E and SST-2 use task-specific answer or sentiment formats.
- Reasoning condition: Without chain-of-thought prompting, models perform slightly worse overall, with the largest rationale-selection difference for Mistral.As a non-reasoning model, Mistral benefits more from chain-of-thought prompting for rationale selection.
C.2 HateXplain Additional Results
Additional HateXplain analyses examine persona agreement, label-distribution shifts, disagreement rates, rationale performance, and chain-of-thought effects across models and demographic groups.
- Label distributions: Significant label-distribution shifts occur despite high inter-persona agreement, especially for Political View and Loneliness.Pairwise Stuart-Maxwell tests with Bonferroni correction assess these within-aspect distributional differences.
- Label distributions: 53% of Mistral-Medium Left-wing predictions were Hate Speech, compared with 32% for Right-wing and 24% for Centrist personas.
- Persona disagreement: Political View produced disagreement rates up to 56.4% for Qwen3-32B, which had higher disagreement across all attributes than the other models.Religion also generated substantial divergence across models, according to the table summary.
- Human-group alignment: The best-performing personas do not always match their corresponding demographic groups, while persona-versus-baseline differences remain slight on CoS-E and SST-2.
- Rationale performance: Rationale IOU-F1 is evaluated for baseline and single-attribute personas, including comparisons across demographic groups with group-specific ground-truth labels.