Source-linked AI summary
Quantifying the Persona Effect in LLM Simulations
Tiancheng Hu, Nigel Collier
TL;DR
Subjective NLP annotations reflect diverse perspectives, but the explanatory power of available persona variables is limited. The paper quantifies this variance and evaluates zero-shot persona prompting for predicting individual annotations. Persona prompting produces modest, statistically significant gains, with effectiveness increasing when persona variables better explain human responses, while weakly explanatory datasets limit the benefits.
Problem
The paper asks how much persona variables explain human annotation variance and whether LLMs can use them to simulate diverse perspectives in subjective NLP tasks.
Method
The study uses mixed-effect regression to quantify persona-explained variance and compares zero-shot LLM predictions with and without persona prompting.
Results
Persona variables explain less than 10% of variance across most considered NLP datasets, while persona prompting provides modest but statistically significant improvements.
Takeaways & Limitations
Persona prompting is most useful when annotators frequently disagree within a relatively narrow range, but its benefits are limited when persona variables have weak explanatory power.
Takeaways & Limitations
The paper’s evidence is largely based on English-language settings, and persona-based simulation may be considerably weaker in non-English languages because language and identity terms are culturally specific.
Abstract
from arXiv · showhide
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.
1 Introduction
The paper examines whether persona information helps LLMs simulate diverse perspectives in subjective NLP tasks, where annotations often lack definitive answers. It finds that persona variables explain limited annotation variance, so prompting yields only bounded benefits.
- Subjective NLP annotations rarely have definitive right or wrong answers and commonly show low inter-annotator agreement.
- The study addresses how much persona variables explain human annotation variance and whether persona prompting improves LLM predictions.
- Persona variables explain less than 10% of annotation variance in the considered NLP datasets, while prompting produces modest but statistically significant improvements.
- Persona variables include demographic, social, attitudinal, behavioral, lived-experience, and values-related characteristics.
- Most existing NLP datasets provide no information about their annotators, limiting the available persona information for simulation.
- Persona prompting prepends an annotator’s persona information before the text sample and task description to simulate diverse perspectives.
2 Related Work
Prior work establishes that persona affects subjective NLP annotations but often studies isolated variables or output diversity. This paper instead evaluates persona prompting as a way to predict task annotations and separates persona effects from text-sample effects.
- Earlier studies link persona variables to annotations in hatefulness, sentiment, irony, and related subjective NLP tasks.
- Existing research often lacks a holistic account of how much persona variables explain annotation variance.
- Some studies use personas to diversify model outputs, whereas others model individual or group annotation differences through persona-specific layers.
- This work studies the understudied potential of LLM persona prompting for simulating different perspectives in NLP tasks.
- Unlike alignment studies comparing survey response distributions, this paper tests whether persona variables inform task predictions.
3 RQ1: How much variance in human annotation could persona variables explain?
The paper estimates persona-related annotation variance with mixed-effect regression while controlling for text-specific variability. Across subjective NLP datasets, persona variables explain only a small share of variance, leaving substantial unexplained variation.
- Methodology: A mixed-effect linear regression estimates persona variables’ fixed effects while controlling for text-specific variability with a random intercept for each text.The study evaluates 10 subjective NLP datasets containing unaggregated annotations and annotator persona variables.
- Results: 1.4%-10.6% of total annotation variance is explained by persona variables after controlling for text variation.These values are reported as Marginal R2 values from the regression models.
- Results: Text-specific variability can explain up to 70% of total variance in the considered analysis, exceeding the contribution of persona variables.
- Caveat: The regression is interpretable but a weak baseline for LLMs because it assumes linear persona effects and omits interactions between persona variables.
- Limitations: 25%-70% of variance remains unexplained by either text or persona variables across the tasks considered.Possible contributors include uncollected personality traits, complex beliefs, lived experiences, and interpersonal dynamics.
- Results: Hatefulness assessment allows more personal interpretation, so persona factors may explain less annotation variance than in more predictable tasks.
- Implication: Regression analysis can provide preliminary expectations for persona-prompted LLM performance when annotated data are available.
4 RQ2: Can incorporating persona variables via prompting improve LLMs’ predictions?
The study compares zero-shot LLM predictions with and without persona information across four subjective NLP datasets. Persona prompting improves performance modestly and significantly overall, but gains vary by dataset and model.
- Methodology: The experiments prepend persona variables in zero-shot prompts and compare predictions with and without personas across four subjective NLP datasets.The evaluation targets individual annotations rather than aggregated labels.
- Methodology: Performance is compared using R2, Cohen’s Kappa, MAE, and macro F1, with bootstrap testing for statistical significance.
- Results: Persona prompting yields statistically significant improvement on at least one metric in each of the four datasets, but the gains are generally modest.
- Results: In EPIC, where persona variables explain up to 9% of annotation variance, persona prompting provides a 1% average gain.
- Results: GPT-4 consistently outperforms the other models across tasks, while persona prompting improves some models but not others.
- Robustness: Persona-variable ordering and language variations produce consistent model performances in the robustness experiments.
5 RQ3: For what types of samples is persona prompting most useful?
Persona prompting helps most when annotators disagree substantially but within a relatively narrow range, whereas consensus or wide disagreement leaves less room for useful adjustment.
- Sample categorization: 600 samples per category were created across four groups defined by low/high annotation entropy and standard deviation.The groups were low entropy-low standard deviation, low entropy-high standard deviation, high entropy-low standard deviation, and high entropy-high standard deviation.
- Prediction changes: Persona information produced only slight prediction changes for samples with low entropy.The experiment was averaged across four models and replicated on the POPQUORN-Politeness dataset.
- Prediction changes: Larger prediction shifts occurred for high-entropy, low-standard-deviation samples, where disagreement was substantial but within a small margin.Such adjustments could move a prediction from 3 without persona variables to 4 with them.
- Prediction changes: High entropy combined with high standard deviation made persona-based adjustment more challenging because predictions would need to move far from the mean.The passage gives a hypothetical shift from 0 without persona variables to 4 with persona variables.
- Statistical results: MAE improvements were significant in all four categories across both datasets, but high entropy-low standard deviation showed the strongest comparative improvement.Bootstrapping used 1,000 runs; category differences were tested with one-way ANOVA and Tukey’s range test.
6 RQ4: How effectively can LLMs simulate personas when the importance of persona variables varies?
The study evaluates persona simulation by varying how predictive persona variables are of human responses and comparing prompted LLM predictions with regression-based targets. Prediction accuracy rises with persona predictability, but even strong 70b models remain below the trained-regression ceiling and are unreliable when marginal R2 is below 0.1.
- Methodology: The case study uses ANES survey questions and fixed persona variables to test simulations across questions with varying predictability.After filtering and downsampling, the sample contains 600 respondents and 21 questions, producing 42 persona-question combinations.
- Evaluation: Figure 3 plots target R2 on the x-axis and predicted R2 on the y-axis for each persona-prompting experiment.The figure includes a fitted regression line and the y = x line representing maximum linear-regression performance.
- Results: Predicted R2 increases with target R2, and Tulu-2-70b-dpo captures 81% of the target R2.Other 70b models show similar capabilities except base Llama-2-70b, while 7b and 13b models perform worse.
- Results: No model surpasses y = x, so persona prompting remains below the performance of a trained linear model.The comparison uses predicted R2 from persona prompting against target R2 derived from the regression-based target.
- Results: When target R2 is low, especially below 0.1, even the best models use persona information poorly; sufficiently predictive personas enable somewhat accurate simulations on large models.The reported relationship includes interaction terms between persona and target variables, suggesting the true target R2 may be higher.
- Implications and limitations: Most existing NLP datasets have marginal R2 < 0.1, leading the authors to argue that persona prompting cannot reliably simulate different perspectives in current NLP tasks.This finding is presented as a limitation of applying persona prompting to existing subjective NLP datasets.
- Implications and limitations: Group-level persona variables may mismatch individual identities, while LLMs may represent groups as monoliths and miss within-group heterogeneity.The authors note that finer-grained group variables might improve individualization, but whether they do so remains unresolved.
7 Conclusion and Recommendation
Persona variables explain less than 10% of annotation variance across most considered NLP datasets, so persona prompting yields only modest improvements. The authors recommend caution, strategic dataset design, and validation when fidelity matters.
- Less than 10% of human-annotation variance is accounted for by persona variables across most considered NLP datasets.
- Persona prompting provides modest but statistically significant improvements across tasks.
- The largest gains occur when annotators substantially disagree but their judgments differ only slightly, corresponding to high entropy and low standard deviation.
- The authors advise caution when using LLMs for simulation where persona variables have weak influence.
- Zero-shot simulations intended to faithfully reproduce human behavior require thorough validation and potentially fine-tuning.
- Dataset persona collection should be driven by clear objectives and use targeted variables when behavioral simulation requires higher predictability.
8 Limitations
The study's conclusions are limited by subjectivity, incomplete cultural and linguistic coverage, potential bias, and constrained prompt engineering. These boundaries restrict how broadly its simulation findings should be generalized.
- Human behavior is subjective and shaped by contextual factors that are difficult to quantify, so some error is likely to persist.
- Most available datasets with annotator persona information come from the U.S., limiting conclusions about other cultural contexts.
- No publicly available datasets were identified with annotator persona variables in languages other than English.
- Non-English simulation may be weaker because LLM performance disparities and culturally rooted identity terms cannot be readily translated.
- Zero-shot simulations may be affected by social biases and stereotypes present in LLM pretraining data.
- The study did not conduct extensive prompt engineering because of computational limitations and its targeted scope.
9 Ethical Considerations
The study uses anonymized public persona data and recognizes risks from categorizing people and applying LLM simulations. Its prompts operationalize demographic, social, behavioral, and attitudinal characteristics in task-specific questions.
- The study uses persona variables from anonymized publicly available datasets, so it involved no human participants or newly collected personal data.
- The authors acknowledge risks including identity fraud, manipulation, and concerns about categorizing individuals in persona datasets.
- The demographic categorizations follow practices such as those used by the U.S. Census Bureau and are intended to study diverse-perspective simulation.
- The experiments use pretrained models including GPT-4, GPT-3.5, Llama-2, Llama-2 Chat, Tulu-2, and Tulu-2-dpo.
- Prompt templates combine profile attributes with questions about offensiveness, toxicity, irony, and politeness.
- A leave-one-out regression experiment estimates each persona variable's importance through its change in R2.
B.3 Results from All Models
Full supplementary results compare LLM performance with and without persona prompting, while persona-variable importance is assessed through leave-one-out regression.
- Smaller 7b and 13b models generally perform worse than larger models with and without persona prompting.
- Persona-variable importance scores are computed as the R2 decrease after removing each variable from an all-variable regression model.
C Supplementary Results for Section 6
Supplementary results document the target variables, persona-prompting plots, and robustness tests used for Section 6. Prompt-variable ordering and paraphrasing produce minimal variation in performance.
- The persona template begins with an “It is 2012. Your Profile” framing and specifies demographic information.
- Figure 4 plots predicted R2 against target R2 for persona-prompting experiments, alongside a fitted regression line and the y = x maximum-performance line.
- Robustness checks vary persona-variable order and paraphrase the prompt template across five configurations each, with minimal resulting variation.
- Table 4 compares LLM annotation prediction with and without persona prompting using R2, Cohen’s Kappa (κ), and MAE.
- Target variables and their associated Target R2 values are listed for the Section 6 experiment.