[Paper Review] Quantifying the Persona Effect in LLM Simulations

Tiancheng Hu, Nigel Collier|arXiv (Cornell University)|Feb 16, 2024

Human-Automation Interaction and SafetyPsychology3 citations

TL;DR

This study quantifies the impact of persona variables—such as demographics, attitudes, and lived experiences—on human annotations in subjective NLP tasks and evaluates their utility in improving LLM-based simulation through persona prompting. It finds that persona variables explain less than 10% of variance in most datasets, limiting the effectiveness of persona prompting, which only modestly improves predictions when variance is higher, and performs poorly when persona utility is low.

ABSTRACT

Large language models (LLMs) have shown remarkable promise in simulating human language and behavior. This study investigates how integrating persona variables-demographic, social, and behavioral factors-impacts LLMs' ability to simulate diverse perspectives. We find that persona variables account for <10% variance in annotations in existing subjective NLP datasets. Nonetheless, incorporating persona variables via prompting in LLMs provides modest but statistically significant improvements. Persona prompting is most effective in samples where many annotators disagree, but their disagreements are relatively minor. Notably, we find a linear relationship in our setting: the stronger the correlation between persona variables and human annotations, the more accurate the LLM predictions are using persona prompting. In a zero-shot setting, a powerful 70b model with persona prompting captures 81% of the annotation variance achievable by linear regression trained on ground truth annotations. However, for most subjective NLP datasets, where persona variables have limited explanatory power, the benefits of persona prompting are limited.

Motivation & Objective

• To assess the extent to which persona variables explain variance in human annotations across subjective NLP tasks.
• To evaluate whether persona prompting improves LLMs’ ability to simulate diverse perspectives in subjective NLP annotations.
• To identify the types of data samples for which persona prompting yields the greatest improvement.
• To examine the linear relationship between the explanatory power of persona variables and the performance gain of LLMs in persona simulation.
• To provide recommendations for dataset design and cautious use of LLMs in simulating human perspectives.

Proposed method

• Measured the proportion of variance in human annotations explained by persona variables using R² statistics across four subjective NLP datasets.
• Applied persona prompting by prepending persona descriptions (e.g., demographics, attitudes) to input prompts before LLM inference.
• Evaluated LLM performance using correlation and R² metrics between model predictions and human annotations.
• Conducted controlled experiments with fixed text randomness and varied persona utilities to isolate simulation capacity.
• Used large, instruction-tuned LLMs (e.g., Llama 3, Mixtral) to assess performance across different model scales.
• Performed robustness checks on prompt phrasing and persona order to ensure consistency in results.

Experimental results

Research questions

• RQ1 RQ1: How much variance in human annotation can persona variables explain?
• RQ2 RQ2: Can incorporating persona variables via prompting improve LLMs’ predictions?
• RQ3 RQ3: For what types of samples is persona prompting most useful?
• RQ4 RQ4: How well can LLMs simulate personas when persona utility varies and text randomness is controlled?

Key findings

• Persona variables explain less than 10% of the variance in human annotations across most subjective NLP datasets.
• LLM performance gains from persona prompting are modest and only occur when persona variables explain a meaningful portion of annotation variance.
• The strongest improvement in LLM predictions occurs on samples with frequent annotator disagreements confined to a narrow range.
• A linear relationship exists between the proportion of variance explained by persona variables and the improvement in LLM predictions.
• Large, fine-tuned LLMs can explain up to 81% of the variance in human responses when persona utility is high, but performance drops to near zero when R² < 0.1.
• Most existing NLP datasets fall into the low-utility category (R² < 0.1), limiting the practical effectiveness of persona prompting in current NLP research.