Population-Aligned Persona Generation for LLM-based Social Simulation

Zhengyu Hu; Jianxun Lian; Zheyuan Xiao; Max Xiong; Tianfu Wang; Teng Xiao; Yuxuan Lei; Fengqing Jiang

cs.CLSep 12, 2025

Population-Aligned Persona Generation for LLM-based Social Simulation

Authors: Zhengyu Hu, Jianxun Lian, Zheyuan Xiao, Max Xiong, Tianfu Wang, Teng Xiao, Yuxuan Lei, Fengqing Jiang,

Organizations: HKUST · Microsoft Research Asia · Duke University · Northwestern University ·

Abstract

Recent advances in large language models (LLMs) have enabled large-scale, high-fidelity social simulations, creating new opportunities for computational social science. However, constructing persona sets that faithfully reflect real-world population diversity remains a key challenge. Existing studies often emphasize agentic frameworks and simulation environments, while paying less attention to persona generation and the biases introduced by unrepresentative persona sets. In this paper, we propose a systematic framework for synthesizing high-quality, population-aligned persona sets for LLM-driven social simulation. Our approach begins by leveraging LLMs to generate narrative personas from long-term social media data, followed by rigorous quality assessment to filter out low-fidelity profiles. We then apply importance sampling to achieve global alignment with reference psychometric distributions, such as the Big Five personality traits. To address the needs of specific simulation contexts, we further introduce a task-specific module that adapts the globally aligned persona set to targeted subpopulations. Extensive experiments demonstrate that our method significantly reduces population-level bias and enables accurate, flexible social simulation for a wide range of research and policy applications.

Figures & tables

Figure 1: Comparison of Big Five personality trait distributions across real human data, LLM responses without personas, existing persona sets, and GPT-4o-generated personas.

Figure 2: Overview of the Population-Aligned Persona Generation Framework.

Model Base		Qwen2.5-72B				Llama-3-70B				Phi-4				Avg.
		AMW	FD	SW	MMD	AMW	FD	SW	MMD	AMW	FD	SW	MMD
Without Persona	DAns (temp=0)	0.4111	1.4256	0.4112	0.8631	0.6469	3.4211	0.6962	2.1904	0.4560	1.7358	0.4602	1.1207	1.1532
	DAns (temp=0.7)	0.3860	1.2309	0.3742	0.6969	0.6271	3.0539	0.6611	1.8805	0.3909	1.3098	0.3960	0.8257	0.9861
	DAns (temp=1)	0.3687	1.0947	0.3560	0.6250	0.5946	2.9176	0.6422	1.7867	0.3731	1.1475	0.3744	0.7294	0.9175
With Persona	SyncP (Qwen2.5-72B)	0.2639	0.5854	0.2792	0.3193	0.3124	0.8279	0.3102	0.4749	0.3284	0.8232	0.3541	0.4282	0.4423
	SyncP (Llama-3-70B)	0.2759	0.5832	0.2679	0.3392	0.3996	0.8307	0.3472	0.4421	0.3336	0.8268	0.3383	0.4886	0.4561

Table 1: In-domain alignment results on the IPIP Big Five dataset for three base models across four metrics Lower is better for all metrics. Best results are red , second-best blue .

	CFCS				FBPS				Duckworth				Avg.
	AMW	FD	SW	MMD	AMW	FD	SW	MMD	AMW	FD	SW	MMD
AlignX	0.2189	0.0697	0.2457	0.1091	0.3527	1.0808	0.3122	0.5710	0.4083	0.4333	1.2429	0.7111	0.4796
Tulu-3-Persona	0.2755	0.0826	0.3086	0.1450	0.2418	0.4849	0.2716	0.2564	0.3838	0.4009	0.7533	0.5815	0.3488
Bavard	0.2415	0.0870	0.2708	0.1171	0.2773	0.6788	0.2545	0.3717	0.3048	0.3052	0.7038	0.4301	0.3369
Google Synthetic	0.2563	0.0972	0.2873	0.1320	0.2857	0.7112	0.2648	0.3985	0.3141	0.3122	0.7295	0.4524	0.3534
Nvidia Nemotron	0.2514	0.1029	0.2794	0.1178	0.2595	0.9599	0.2817	0.5387	0.2970	0.4054	1.1136	0.6837	0.4409

Table 2: Out-of-domain alignment results on CFCS, FBPS, and Duckworth datasets across four metrics. Lower is better for all metrics. Best results are red , second-best blue .

Method	BIG5	CFCS	FBPS	Duwo	Avg.
AlignX	0.5761	0.7944	0.5078	0.5688	0.6118
Tulu-3-Persona	0.4971	1.0521	0.4017	0.4970	0.6120
Bavard	0.4528	0.4066	0.4268	0.4559	0.4355
Google Synthetic	0.4570	0.4382	0.4168	0.4571	0.4423
Nvidia Nemotron	0.8000	1.1935	0.6894	0.8075	0.8726
SyncP (GPT-4o)	1.0412	1.2794	0.9289	1.0443	1.0735

Table 3: MAE corr between human and synthetic inter-trait correlations. Lower values indicate better individual-level consistency. Best results are red , second-best blue .

	YRBSS				WVS (East Asia)				WVS (North America)				WVS (Western Europe)				Avg.
	AMW	FD	SW	ED	AMW	FD	SW	ED	AMW	FD	SW	ED	AMW	FD	SW	ED
AlignX	0.4676	9.7132	0.5679	2.0409	0.8274	21.0512	0.8712	2.4598	0.8785	20.4461	0.8914	2.4495	0.8595	23.4058	0.8576	2.0302	5.6136
Tulu-3-Persona	0.4921	9.0356	0.7252	1.9259	0.9282	23.2265	0.9343	2.6507	0.8498	20.8217	0.8976	2.6438	0.9485	26.8801	0.9615	2.3601	6.0176
Bavard	0.4569	9.0103	0.5558	1.8776	0.8153	22.7143	0.8190	2.1368	0.8489	20.3607	0.8730	2.4311	0.8991	24.8798	0.8593	1.9787	5.7198
Google Synthetic	0.4716	9.0975	0.5500	1.8997	0.8197	21.5079	0.8443	2.2662	0.8412	22.6565	0.8757	2.6811	0.8818	24.8154	0.8667	1.9872	5.8164
Nvidia Nemotron	0.5052	10.9920	0.6999	2.2383	0.8845	24.7850	0.8990	2.6935	0.8687	20.2640	0.8733	2.4560	0.8793	26.9882	0.8832	2.1464	6.1910

Table 4: Performance of group-specific persona generation across datasets and regions. ED denotes Energy Distance. Lower is better for all metrics. Best results are red , second-best blue .

Figure 3: Scatter plots comparing the distribution of EXT and EST scores between real human responses (blue) and LLM-generated outputs (orange) across nine settings.

Figure 4: Result for social science experiments. Higher is better for Accuracy Score; lower is better for Distribution Score.

Setting	RandomSelect	Ours	Nvidia	AlignX	Google
KDE bandwidth $h$ (default: $0.20$ )
$h=0.10$	0.22	0.17	0.26	0.34	0.31
$\mathbf{h=0.20}$	0.22	0.16	0.26	0.34	0.31
$h=0.30$	0.22	0.16	0.26	0.34	0.31
$h=0.50$	0.22	0.17	0.26	0.34	0.31
IS Top- $k$ (default: $70\%$ )

Table 6: Sensitivity analysis of key hyperparameters in the two-stage sampler on IPIP Big Five (AMW; lower is better) with Qwen2.5-72B. We vary one parameter at a time and keep others at defaults (

h{=}0.20

, Top-

k{=}70\%

\varepsilon{=}0.08

, Sinkhorn iters

T{=}250

, persona size

N^{\prime}{=}5\mathrm{k}

). Numbers are reported as (RandomSelect / Ours / Nvidia / AlignX / Google).

Appendix figures & tables21 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	# Questions	# Human Responses	# Countries and Regions
Big Five (IPIP)	50	1,015,341	223
CFCS	12	15,034	144
FBPS	76	41,841	169
Duckworth	12	4,270	111
YRBSS (2023)	107	20,103	1
WVS (Full)	294	97,221	66

Appendix

Table 9: Statistics of datasets used for alignment and evaluation.

Test	Reference	AMW	FD	SW	MMD
CFCS	FBPS	0.1976	0.0685	0.1974	0.0715
	Duckworth	0.1990	0.0725	0.1990	0.0713
	IPIP	0.1954	0.0614	0.1726	0.0642
FBPS	CFCS	0.2340	0.4653	0.2080	0.2400
	Duckworth	0.2442	0.4750	0.2050	0.2550
	IPIP	0.2223	0.4220	0.1900	0.2186

Appendix

Table 11: Cross-instrument reference ablation with different alignment references. Each block evaluates transfer to a held-out psychometric instrument. Lower is better for all metrics.

	Human		GPT-5.5
WVS setting	No leakage	Potential	No leakage	Potential
Hong Kong	68/70	2/70	66/70	4/70
Germany	67/70	3/70	65/70	5/70
United States	66/70	4/70	67/70	3/70
Overall	201/210	9/210	198/210	12/210

Appendix

Table 12: Stereotype-leakage audit of group-specific adaptation. No leakage denotes a score of at least 4; Potential denotes a score below 4.

Method	AMW	FD
Full pipeline	0.2215	0.4630
w/o text rewriting	0.3231	0.7834
w/o persona quality filter	0.2673	0.5726

Appendix

Table 13: Ablation study of the seed persona mining pipeline on IPIP Big Five with Qwen2.5-72B. Lower is better for both metrics.

Sampler	AMW	FD
IPF [ 67 ]	0.2063	0.3675
IDS [ 67 ]	0.1985	0.3446
Ours (IS)	0.1834	0.3241
Ours (IS+OT)	0.1619	0.2858

Appendix

Table 14: Comparison with SocioVerse sampling baselines on IPIP Big Five under the Qwen2.5-72B setting.

Setting	$N^{\prime}{=}1\mathrm{k}$	$N^{\prime}{=}5\mathrm{k}$	$N^{\prime}{=}10\mathrm{k}$	$N^{\prime}{=}80\mathrm{k}$
Aligned pool size $N^{\prime}$ (default: $5\mathrm{k}$ )
YRBSS	0.433	0.290	0.287	0.285
WVS (East Asia)	0.896	0.784	0.782	0.781
WVS (North America)	0.822	0.774	0.771	0.770
WVS (Western Europe)	0.878	0.790	0.786	0.783

Appendix

Table 15: Aligned pool size sensitivity on downstream subgroup alignment (AMW; lower is better).

Game	Ours	Nvidia	AlignX	Google
Trust	0.05	0.08	0.07	0.09
Dictator	1.53	2.23	1.75	2.19
MAP Trust	0.11	0.26	0.19	0.22
Risky Dictator	0.22	0.26	0.24	0.25
LPG	0.12	0.24	0.27	0.26
LGG	0.19	0.39	0.31	0.40

Appendix

Table 16: Behavioral generalization on six tasks (deviation from human reference; lower is better).

Persona source	Action Accuracy	Session Outcome F1
AlignX	20.72	51.86
Bavard	22.08	54.73
Google Synthetic	21.91	54.25
Nvidia Nemotron	20.54	51.29
Ours (Resample)	23.84	56.61

Appendix

Table 17: Multi-turn customer behavior simulation. Higher is better.

Source	Raw	Preproc (S1)	Raw QC (S2)	Persona Gen (S3)	Persona QC (S4)
Reddit	47,200,000	4,000,000	2,800,000	102,861	75,661
Twitter	88,100,000	1,700,000	673,000	75,243	54,010
Blog	681,000	500,000	368,000	41,607	30,738
Total	135,981,000	6,200,000	3,841,000	219,711	160,409

Appendix

Table 18: Seed-persona data statistics. Counts in Raw, Raw QC, Persona Gen, and Persona QC are from HumanLLM [ 28 ] (Table A1).

Stage	Evaluator	Faithful ( $\geq 4$ )	Potential shift ( $<4$ )
Post rewriting	Human	207/210	3/210
Post rewriting	GPT-5.5	208/210	2/210
Persona generation	Human	208/210	2/210
Persona generation	GPT-5.5	208/210	2/210

Appendix

Table 19: Source-faithfulness audit. Each stage includes 210 pairs; scores below 4 indicate a potential content shift.

Method	IPIP	CFCS	FBPS	Duckworth
Tulu-3-Persona	$83.83\pm 1.37$	$8.33\pm 0.67$	$48.59\pm 0.76$	$40.01\pm 0.59$
Bavard	$78.26\pm 0.94$	$8.76\pm 0.58$	$68.01\pm 0.98$	$30.45\pm 0.48$
Google Synthetic	$80.68\pm 1.04$	$9.78\pm 0.58$	$71.26\pm 1.06$	$31.14\pm 0.59$
Nvidia Nemotron	$83.00\pm 1.28$	$10.34\pm 0.43$	$96.12\pm 0.98$	$40.45\pm 0.62$
Ours (Raw)	$46.21\pm 0.64$	$7.02\pm 0.20$	$46.51\pm 0.64$	$23.26\pm 0.32$
Ours (RandomSelect)	$46.09\pm 0.80$	$6.99\pm 0.25$	$46.35\pm 0.80$	$23.19\pm 0.40$

Appendix

Table 20: FD across five runs (

\times 100

; lower is better). IPIP is the alignment reference; the other instruments are held out.

Dataset	Metric	RandomSelect	Resample	$p$ -value
CFCS	AMW	$19.69\pm 0.27$	$19.50\pm 0.32$	0.0103
	FD	$6.99\pm 0.25$	$6.18\pm 0.30$	0.0018
	$\mathrm{MAE}_{\mathrm{corr}}$	$24.67\pm 0.41$	$22.94\pm 0.46$	0.0003
FBPS	AMW	$23.67\pm 0.32$	$22.18\pm 0.36$	0.0001
	FD	$46.35\pm 0.80$	$42.32\pm 0.87$	0.0001
	$\mathrm{MAE}_{\mathrm{corr}}$	$38.87\pm 0.58$	$36.75\pm 0.66$	0.0007

Appendix

Table 21: OOD comparisons across five runs (scores

\times 100

; lower is better).

Method	Accuracy $\uparrow$	Distribution $\downarrow$
Google Synthetic	$68.4\pm 0.52$	$23.6\pm 0.39$
Nvidia Nemotron	$69.5\pm 0.61$	$22.5\pm 0.34$
Ours (Raw)	$71.7\pm 0.28$	$21.2\pm 0.19$
Ours (RandomSelect)	$71.6\pm 0.36$	$21.2\pm 0.24$
Ours (Resample)	$72.3\pm 0.31$	$21.0\pm 0.18$
$p$ -value	0.0001	0.0403

Appendix

Table 22: SocSci results across five runs (scores

\times 100

). The last row compares RandomSelect with Resample.

Persona collection	Selection	CFCS	FBPS	Duckworth
Bavard	Random	$8.76\pm 0.58$	$68.01\pm 0.98$	$30.45\pm 0.48$
Bavard	IS+OT	$7.64\pm 0.62$	$62.10\pm 1.10$	$27.42\pm 0.52$
Google Synthetic	Random	$9.78\pm 0.58$	$71.26\pm 1.06$	$31.14\pm 0.59$
Google Synthetic	IS+OT	$8.25\pm 0.61$	$65.06\pm 1.15$	$28.11\pm 0.63$
Nvidia Nemotron	Random	$10.34\pm 0.43$	$96.12\pm 0.98$	$40.45\pm 0.62$
Nvidia Nemotron	IS+OT	$8.64\pm 0.48$	$87.76\pm 1.12$	$35.42\pm 0.68$

Appendix

Table 23: IS+OT on external collections: held-out psychometric FD (

\times 100

; lower is better).

Persona collection	Selection	East Asia	North America	Western Europe
Bavard	Random	$2269.17\pm 24.13$	$2033.71\pm 21.17$	$2490.76\pm 31.00$
Bavard	IS+OT	$2155.71\pm 27.10$	$1962.02\pm 24.06$	$2326.22\pm 34.18$
Google Synthetic	Random	$2147.76\pm 32.38$	$2262.66\pm 26.83$	$2484.93\pm 37.82$
Google Synthetic	IS+OT	$2040.37\pm 35.02$	$2149.53\pm 29.14$	$2320.68\pm 40.31$
Nvidia Nemotron	Random	$2474.93\pm 38.12$	$2023.15\pm 29.24$	$2702.71\pm 43.42$
Nvidia Nemotron	IS+OT	$2221.18\pm 41.05$	$1921.99\pm 32.06$	$2447.57\pm 46.10$

Appendix

Table 24: IS+OT on external collections: WVS FD (

\times 100

; lower is better).

Persona collection	Selection	Accuracy $\uparrow$	Distribution $\downarrow$
Bavard	Random	$68.5\pm 0.46$	$22.7\pm 0.31$
Bavard	IS+OT	$70.8\pm 0.41$	$21.8\pm 0.27$
Google Synthetic	Random	$68.4\pm 0.52$	$23.6\pm 0.39$
Google Synthetic	IS+OT	$70.3\pm 0.34$	$22.1\pm 0.15$
Nvidia Nemotron	Random	$69.5\pm 0.61$	$22.5\pm 0.34$
Nvidia Nemotron	IS+OT	$71.2\pm 0.45$	$21.8\pm 0.38$

Appendix

Table 25: IS+OT on external collections: SocSci scores (

\times 100

Method	AMW	FD	SW	ED
East Asia: Male
AlignX	0.8957	23.0869	0.9428	2.6813
Bavard	0.8792	24.8543	0.8826	2.3175
Google Synthetic	0.8904	23.6874	0.9154	2.4814
Nvidia Nemotron	0.9635	27.4403	0.9802	2.9617
PersonalHub	0.8811	25.1228	0.8946	2.4056

Appendix

Table 26: Gender subgroup evaluation on WVS. ED denotes Energy Distance. Lower is better for all metrics.

Method	IPIP	CFCS	FBPS	Duckworth
SYNTHIA	0.3516	0.1168	0.3595	0.4372
Ours	0.0885	0.0642	0.2186	0.2623

Appendix

Table 27: Comparison with SYNTHIA using Qwen2.5-72B. Lower MMD is better.

Method	Frobenius norm	$\alpha$ gap
AlignX	11.98	0.0852
Bavard	12.35	0.0879
Google Synthetic	12.53	0.0891
Ours	9.80	0.0697

Appendix

Table 28: Internal response structure relative to humans. Lower is better.

Method	Original	Paraphrased	Reversed
Google Synthetic	$80.68\pm 1.04$	$81.35\pm 1.09$	$82.16\pm 1.16$
Nvidia Nemotron	$83.00\pm 1.28$	$82.71\pm 1.34$	$84.54\pm 1.41$
Ours (Resample)	$28.50\pm 0.59$	$28.02\pm 0.64$	$29.71\pm 0.70$

Appendix

Table 29: IPIP FD under prompt variation (mean

\pm

standard deviation,

\times 100

; lower is better).

Response model	AMW	FD	SW	MMD
Qwen2.5-72B	0.2645	0.8316	0.3199	0.4414
GPT-5.5	0.2481	0.7014	0.2617	0.3572

Appendix

Table 30: IPIP evaluation of Nvidia Nemotron personas with two response models. Lower is better.

Large Language Model-Based Social Simulation

Large language model-based social simulation studies how language-model agents represent people and interact to produce social behavior.

See all 282 papers

MetaPersona: Task-Grounded Synthetic Populations from Empirical Social Science
Sep 29, 2026Jinyi Ye, Yuangang Li, Chenxiao Yu +7Personality Modeling in Language Models Synthetic Data Generation
University of Southern California · University of California, Irvine · 2077AI +1
How Well Do Large Language Models Capture Human Personality?
May 12, 2026Aanisha Bhattacharyya, Yaman Kumar Singla, Rajiv Ratn Shah +2Persona-Conditioned Generation Representation Collapse
Adobe Media and Data Science Research (MDSR) · IIIT-Delhi · SUNY at Buffalo
Persona-Based Simulation of Human Opinion at Population Scale
Mar 28, 2026Mao Li, Frederick G. ConradHuman Behavior Simulation Large Language Model-Based Social Simulation
Institute for Social Research, University of Michigan, 426 Thompson St., Ann Arbor, 48104, MI, US.

Sep 29, 2026J/K move · Enter open · S save

MetaPersona: Task-Grounded Synthetic Populations from Empirical Social Science

Jinyi Ye, Yuangang Li, Chenxiao Yu, Preyashi Poddar+6

University of Southern California · University of California, Irvine · 2077AI+1

Personality Modeling in Language Models Synthetic Data Generation Computational Social Science Synthetic Population Generation Large Language Model-Based Social Simulation

Open paper AskPDF arXiv

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.

Abstract

Large Language Model-Based Social Simulation

Explore similar work

MetaPersona: Task-Grounded Synthetic Populations from Empirical Social Science

How Well Do Large Language Models Capture Human Personality?

Persona-Based Simulation of Human Opinion at Population Scale