MetaPersona: Task-Grounded Synthetic Populations from Empirical Social Science

Jinyi Ye; Yuangang Li; Chenxiao Yu; Preyashi Poddar; Priyanka Dey; Longtian Ye; Zihan Wang; Xiyang Hu

cs.AISep 29, 2026

MetaPersona: Task-Grounded Synthetic Populations from Empirical Social Science

Authors: Jinyi Ye, Yuangang Li, Chenxiao Yu, Preyashi Poddar, Priyanka Dey, Longtian Ye, Zihan Wang, Xiyang Hu, +2 more

Organizations: University of Southern California · University of California, Irvine · 2077AI · Arizona State University

Abstract

Personas used to seed LLM social simulations face a cold-start problem: existing methods lack a principled basis for deciding which attributes to include and how to assign their values. As a result, synthetic populations may misrepresent the demographic composition, latent attributes, and dependency structure that shape downstream behavior. We introduce MetaPersona-DB, a dataset of 11,000+ empirical human-subjects studies annotated with task-relevant variables, reported relationships, and aggregate-level population statistics. Building on this resource, we propose MetaPersona, a framework that retrieves task-relevant evidence, constructs literature-derived persona dependency graphs, and samples synthetic populations from empirical priors linking demographics, latent attributes, and outcomes. Across three downstream case studies, three baselines, and three frontier models, results vary by task and model: MetaPersona performs strongly on misinformation belief and AI-tool sentiment, while results on income redistribution are mixed. It also reduces persona-construction cost to under $0.5 per task using GPT-5.2. Finally, we present MetaPersona-Studio, a prototype interactive interface for empirically grounded persona generation.

Figures & tables

Figure 1: The persona cold-start problem and our response. We address the problem through three contributions: MetaPersona-DB , a curated empirical evidence base for persona variables; MetaPersona , a framework for selecting task-relevant attributes and sampling empirically grounded synthetic populations; and MetaPersona-Studio , an accessible low-cost interface for interdisciplinary use.

Figure 2: An example of an annotated paper.

Figure 3: Overview of the MetaPersona pipeline: from a task query to a synthetic population.

Task	Discipline	Dataset	Predicted Outcome Variable
Misinformation belief	Communication, political science	MIST ( Maertens et al., 2024 )	Ability to distinguish real from fake headlines, measured by numerical MIST-20 score
Income redistribution attitude	Economics, sociology	ISSP 2019 Social Inequality V (ZA7600, v3.0.0)	Agreement that government should reduce income differences, measured by ordinal scale
AI-tool sentiment & adoption	Behavioral science, HCI	Stack Overflow Survey ( Stack Overflow, 2025 )	Sentiment toward AI tools and expected future use, measured by ordinal scales

Table 1: Downstream evaluation tasks, ground-truth data sources, and predicted outcomes.

Figure 4: Examples of the four persona-construction methods evaluated for misinformation-belief prediction: Demo-Only , Backstory , LLM-Heuristic , and MetaPersona . Evaluation uses matched responses from real MIST survey participants ( $N=1{,}000$ ) from Maertens et al. (2024) .

		Case Study 1: Misinformation (MIST)					Case Study 2: Redistribution (ISSP)				Case Study 3: AI Tool Sentiment (SO)
Backbone	Method	RMSE $\downarrow$	MAE $\downarrow$	Bias	W-dist $\downarrow$	KS $\downarrow$	MAE $\downarrow$	W-1 Acc $\uparrow$	QWK $\uparrow$	W-dist $\downarrow$	MAE $\downarrow$	W-1 Acc $\uparrow$	QWK $\uparrow$	W-dist $\downarrow$
GPT-5.2	Demo-Only	0.310	0.257	0.248	0.248	0.619	0.812	0.817	0.086	0.583	0.248	0.743	0.059	0.808
	Backstory	0.352	0.304	0.303	0.303	0.896	0.828	0.819	0.111	0.627	0.250	0.749	0.036	0.863
	LLM-Heuristic	0.322	0.268	0.255	0.257	0.666	0.842	0.808	0.128	0.570	0.252	0.735	0.073	0.871
	MetaPersona	0.270	0.216	0.118	0.118	0.346	0.832	0.814	0.087	0.642	0.217	0.743	0.320	0.616
Claude- Haiku 4.5	Demo-Only	0.238	0.192	0.151	0.153	0.475	0.884	0.802	0.170	0.416	0.293	0.667	0.052	0.931

Table 2: Performance of all methods across three case studies and three LLM backbones. Within each base LLM, we compare four persona-construction methods. Best result per task and metric in bold ; second-best is underlined .

\downarrow

= lower is better;

\uparrow

= higher is better. Individual accuracy metrics and distributional alignment metrics are highlighted in the header. Bias is signed mean error; best/second-best for Bias are based on smallest absolute bias. MIST metrics use scores normalized to

[0,1]

; ISSP MAE and W-dist use the original 1–5 ordinal codes. Metric magnitudes should therefore not be compared across tasks. For Case Study 3, W-dist is macro-averaged over two predicted outcomes.

Task / Path	Method	$r$	$\|\Delta r\|\,\downarrow$
MIST CRT $\to$ discern. (expect $+$ )	Ground truth	0.313	—
	LLM-Heuristic	0.238	0.075
	MetaPersona	0.293	0.020
ISSP education $\to$ support (expect $\approx 0$ /mixed)	Ground truth	0.043	—
	Demo-Only	0.107	0.064
	Backstory	0.147	0.104

Table 3: Structural realism using GPT-5.2 ; results for other models are reported in Appendix E . Pearson

r

along one diagnostic path per task in the ground-truth data versus each method’s prediction. Closest to ground truth in bold ;

|\Delta r|

is the absolute deviation from the ground-truth

r

(lower is better).

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Method	MIST	ISSP	SO
Number of personas	1000	2915	2000
LLM-Heuristic	$\sim$ $2.27	$\sim$ $7.91	$\sim$ $4.64
Backstory	$\sim$ $8.52	$\sim$ $27.68	$\sim$ $16.84
MetaPersona	$\sim$ $0.16	$\sim$ $0.25	$\sim$ $0.18

Appendix

Table 4: Persona construction cost across case studies.

Validation component	Measure	Result
Validation 1: Empirical paper labeling
Human agreement	Cohen’s $\kappa$	0.94
Human agreement	Raw agreement	97%
LLM vs. adjudicated ground truth	Precision	86%
LLM vs. adjudicated ground truth	Recall	97%
LLM vs. adjudicated ground truth	F1	91%

Appendix

Table 5: Human validation results. Validation 1 measures empirical-paper labeling reliability and LLM classification accuracy against adjudicated human labels. Validation 2 measures whether extracted annotations are supported by full-text PDFs.

Figure 5: Distribution of articles in MetaPersona-DB across disciplines and journals.

Field	Definition	Examples / Allowed Values
Study-level descriptors
Domain or subfield	The specific empirical or disciplinary area of the study.	Political communication; health psychology; labor economics; organizational behavior.
Agent types	The social actors involved in the study.	Voters; consumers; patients; employees; managers; firms; institutions.
Context: country	The geographic setting of the study.	USA; Germany; China; cross-national sample.
Context: population	The sample, population, or empirical setting studied.	Registered voters; undergraduate students; low-income households; Fortune 500 employees.
Main finding	A one-sentence summary of the paper’s primary empirical result.	“Economic insecurity is associated with lower institutional trust.”

Appendix

Table 6: Annotation taxonomy for extracting persona-relevant variables and variable relationships from paper metadata.

Task / Path	Model	Method	$r$	$\|\Delta r\|\,\downarrow$
MIST CRT $\to$ discern. (expect $+$ )	Claude Haiku 4.5	Ground truth	0.313	—
		LLM-Heuristic	-0.172	0.485
		MetaPersona	-0.021	0.334
	DeepSeek v3.2	Ground truth	0.313	—
		LLM-Heuristic	0.423	0.110
		MetaPersona	0.052	0.261

Appendix

Table 7: Structural realism for additional models. Pearson

r

along one diagnostic path per task in the ground-truth data versus each method’s prediction. Closest to ground truth within each model is shown in bold ;

|\Delta r|

is the absolute deviation from the ground-truth

r

(lower is better).

Explore similar work

ProfileFoundry: A Synthetic Person-Object Substrate for Privacy, Memory, and Tool-Use Evaluation in LLM Agent
Jun 24, 2026Sriram Selvam, Anneswa GhoshProfile Synthetic Data
Beyond Individual Personas: Aligning Synthetic Dialogue to Population-Level Behavior Distributions
Jun 5, 2026Xinyi Liu, Rinat Khaziev, Hooshang Nayyeri +3Conversational Datasets Dialogue
1Amazon · University of Illinois Urbana–Champaign · University College London
How Well Do Large Language Models Capture Human Personality?
May 12, 2026Aanisha Bhattacharyya, Yaman Kumar Singla, Rajiv Ratn Shah +2Artificial Intelligence Personas Persona Consistency
Adobe Media and Data Science Research (MDSR) · IIIT-Delhi · SUNY at Buffalo

Jun 24, 2026, cs.CLJ/K move · Enter open · S save

ProfileFoundry: A Synthetic Person-Object Substrate for Privacy, Memory, and Tool-Use Evaluation in LLM Agent

Sriram Selvam, Anneswa Ghosh

Profile Synthetic Data Generation Provenance User Simulation Population Dynamics Evaluation Agent

Open paper AskPDF arXiv

Foundation-model research increasingly needs data about people: user state, personal histories, relationships, contact-like fields, documents, and longitudinal updates. Real user data is difficult to share, perturb, audit, or redistribute responsibly, while independently generated fake fields rarely preserve the cross-field and temporal consistency needed for controlled evaluation. We present PROFILEFOUNDRY, a deterministic generator and fixed reference release of 100,000 adult synthetic Person Objects across eight locales. Each object combines a typed current snapshot, household, family, and employer links, snapshot-aligned events, normalized relational views, and generation provenance. The release contains 709,228 events, 40,338 households, 52,491 employers, and 518,564 directed relationship edges. We report evidence in separate categories: selected population-marginal comparisons, per-object invariant checks, release-wide referential and temporal closure, and coincidence/provenance screens. PROFILEFOUNDRY is not a population-fidelity model, a rendered-text corpus, or a formal privacy mechanism. Instead, it is a responsible synthetic source layer for constructing downstream foundation-model evaluations involving memory, privacy, document understanding, record linkage, and agent state while keeping the synthetic person behind each artifact inspectable

Abstract

Explore similar work

ProfileFoundry: A Synthetic Person-Object Substrate for Privacy, Memory, and Tool-Use Evaluation in LLM Agent

Beyond Individual Personas: Aligning Synthetic Dialogue to Population-Level Behavior Distributions

How Well Do Large Language Models Capture Human Personality?