Prompts to Proxies: Emulating Human Preferences via a Compact LLM Ensemble

Bingchen Wang; Zi-Yu Khoo; Jingtan Wang

cs.AISep 14, 2025

Prompts to Proxies: Emulating Human Preferences via a Compact LLM Ensemble

Authors: Bingchen Wang, Zi-Yu Khoo, Jingtan Wang

Organizations: Independent Researcher, China · AI Singapore, Singapore · School of Computing, National University of Singapore, Singapore

Abstract

Large language models are increasingly used as proxies for human subjects in social science research, yet external validity requires matching the response distributions of target human populations. We study population-level survey alignment: reconstructing aggregate survey responses from limited public survey data, without individual-level demographic profiles or model finetuning. We formalize this problem as preference reconstruction: rather than matching proxy agents to demographic profiles, we construct a functional basis of proxy agents and recover population preferences by weighted aggregation. We instantiate this idea via Prompts to Proxies ( $\texttt{P2P}$ ), a two-stage inference-time system. Stage 1 uses structured attribute-based prompting with entropy-guided adaptive sampling to construct a diverse proxy pool spanning the latent preference space. Stage 2 employs L1-regularized regression to select a compact weighted ensemble matching observed target-population responses. Across 14 American Trends Panel waves, $\texttt{P2P}$ achieves an average test MSE of 0.014 at approximately 0.8 USD per survey, improving over prompting and demographic-conditioning baselines. On the World Values Survey, cross-locale transfer experiments show that basis expressiveness can outperform locale matching on average, while locale-specific generation helps on culturally divergent questions. A stress test against an SFT-aligned survey model shows competitive performance using less than 3% of the training data. These results position preference reconstruction as a lightweight, externally verifiable alternative for survey-based population alignment. Code is available at https://github.com/bingchen-wang/P2P.

Figures & tables

Figure 1: Left : Finding a functional basis to recover latent ground-truth preference. Right : Identification of latent preference through multiple observations.

Figure 2: System Overview of the Prompts to Proxies ( P2P ) Framework. Stage 1 constructs a response-expressive proxy pool using a structured attribute bank and entropy-guided adaptive sampling strategies. Stage 2 assigns weights to reconstruct population-level preference patterns. Held-out questions afford an external validity check for out-of-sample distributional fidelity.

Wave	Topic	# Qs	MSE	Ent.	Cost	MSE	Ent.	Cost *	MSE	Ent.	Cost	$\Delta$
			Vanilla			PERSONA			P2P
W26	Guns	78	.015	.50	$0.5	.024	.37	$1.0	$\mathbf{.010}$	.55	$0.6	36%
W27	Auto	96	.032	.46	$0.6	.041	.40	$1.3	$\mathbf{.022}$	.54	$0.7	33%
W29	Gender	77	.026	.45	$0.5	.030	.43	$1.0	$\mathbf{.016}$	.51	$0.6	41%
W32	Comm.	98	.019	.55	$0.6	.025	.56	$1.3	$\mathbf{.016}$	.62	$0.7	17%
W34	Biomed	67	.021	.44	$0.5	.018	.39	$0.9	$\mathbf{.011}$	.54	$0.5	39%

Table 1: Performance comparison across 14 ATP waves. Test MSE, entropy, and cost (USD) reported, as means over 3 runs for all methods (sample SDs in Appendix D ).

\Delta

shows relative test MSE improvement over the best baseline; its average is the mean of the per-wave improvements.

Metric	US	GB	HK
Entropy $\uparrow$	.572 $\pm$ .015	.571 $\pm$ .004	.548 $\pm$ .021
Test MSE $\downarrow$	.015 $\pm$ .001	.014 $\pm$ .001	.024 $\pm$ .001
Cost $\downarrow$	1.76 $\pm$ .01	1.71 $\pm$ .00	1.74 $\pm$ .01

Table 2: P2P performance on WVS Wave 7 across three locales. Values are mean

\pm

std over 3 runs. (Full details in Appendix D .)

Figure 3: Basis Expressiveness vs. Locale Matching on WVS HK. We compare US-native and HK-native with HK-transfer (US endowments applied to HK via P2P Stage 2). (A) US and HK ground-truth response patterns are correlated but distinct (

r=0.848

). (B) HK-transfer outperforms HK-native in overall performance, corroborating the functional basis perspective. (C) Error gap scales with US-HK ground-truth divergence, suggesting learnability limits (Definition 2.2 ). (D) Despite higher test MSE, HK-native wins more often on culturally divergent questions, indicating that localization helps in tail cases.

Appendix figures & tables34 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 4: Attribute Bank and Endowment Generation Workflow. The preset attribute bank is organized hierarchically by template (core, thematic, theoretical), mode (e.g., economics, politics, maslow, big five), and attribute (e.g., race, gender, ethnicity, urban/rural residency, risk aversion). Endowment is generated by a specialized LLM ( EndowmentModel ), which takes drawn attributes as input and outputs a profile with instantiation of these attributes.

Figure 5: Left: Entropy change across update steps for each question during active endowment generation for ATP W42. Each column represents a question. Green indicates an increase in entropy from the previous step, while red indicates a decline. Right: Entropy trajectories by question (line chart). Active endowment generation progressively increases response diversity. Trends are classified based on slope and volatility of the entropy trajectory: “rising” if the slope exceeds 0.1 and standard deviation is above 0.02; “falling” if the slope is below –0.1 with sufficient volatility; otherwise labeled “stable”.

Figure 6: Question entropies at the end of active endowment generation.

Figure 7: Cross-validated MSE by alpha ( $\alpha$ ), the penalization parameter, for constrained lasso. The best alpha selected by CV is $1.62\times 10^{-3}$ .

Figure 8: Prompts to Proxies: Emulation Results for ATP W42. Left : Train (left) and test (right) prediction performances for the binarized responses using constrained lasso. Right : Human and agent (pre- and post- alignment) aggregated responses for the sample test question RQ4_F1Ae_W42 .

Figure 9: Breakdown of Selected Agents (58 in total) for W42 by Template. Preset, freeform, and mixed templates all contribute non-trivially to the final ensemble.

Figure 10: Examples of endowments generated by P2P for a run on ATP W42.

Figure 11: Left: Cross-validated MSE by alpha, the overall penalization parameter, and L1 ratio, the penalization weight for the L1 term, for constrained elastic net. The lowest cross-validated MSE is achieved at $\alpha=6.95\times 10^{-3}$ , with l1 ratio set at $0.10$ . Right: Prediction performance of the agent ensemble (107 selected agents) using constrained elastic net. Results are similar to those using constrained lasso.

Figure 12: Mean Squared Error (Test Set) and Precision by Different Observation-to-Agent Ratio (OAR) for Constrained Lasso. Each simulation is repeated 10 times with different random seeds. Subsampling of endowments is done with fixed ground-truth-to-proxy ratio. Subsampling of responses is done with fixed train-valid-test split and ratio.

Figure 13: Snapshots of Weight-vs-Coefficient Comparison and Prediction Performance. Each row represents a simulation run. The red bars on the left panel represent ground truth agent weights used to generate the aggregate data and the blue bars Lasso coefficients. The observation-to-agent ratios for the top, middle and bottom panels are: 0.617, 1.203, 2.407. Lasso’s ability to recover the ground truth agents from the agent pool degrades with lower observation-to-agent ratio, but its predictive accuracy only drops mildly.

Figure 14: Left: Mean Squared Errors and Coefficients of Determination ( $R^{2}$ ) by Number of Proxy Agents. Right: Mean Squared Errors and Coefficients of Determination ( $R^{2}$ ) by Number of Observation. Each simulation is repeated 10 times with different random seeds.

Figure 15: Snapshots of Prediction Performance for Simulation Study 2.

Figure 16: Entropy Coverage Ratio Diagnostics for Simulation Study 2.

Figure 17: Endowment Count per Mode (Sorted by Entropy).

Figure 18: Mean Squared Error and Coefficient of Determination ( $R^{2}$ ) by Entropy Coverage Ratio (ECR) for the trainval set (left) and test test (right). Entropy Coverage Ratio is defined as the ratio of the average response entropies of the proxies and ground-truth agents. LOWESS curves are fit using 10 repetitions.

Figure 19: Snapshots of Prediction Performance for Simulation Study 3.

	Vanilla			PERSONA			P2P
Wave	MSE	1-JSD	EMD	MSE	1-JSD	EMD	MSE	1-JSD	EMD
W26	.015	.820	.072	.024	.772	.082	.010	.868	.056
	(.002)	(.011)	(.004)	(.003)	(.009)	(.004)	(.000)	(.001)	(.001)
W27	.032	.795	.086	.041	.728	.104	.022	.823	.071
	(.004)	(.008)	(.007)	(.001)	(.004)	(.001)	(.001)	(.006)	(.003)
W29	.026	.767	.095	.030	.748	.106	.016	.815	.083

Appendix

Table 3: Full distribution metrics comparison across 14 ATP waves. MSE measures mean squared error of predicted response probabilities. 1-JSD (1 minus Jensen-Shannon Divergence) measures distributional similarity (higher is better). EMD (Earth Mover’s Distance) measures distributional distance (lower is better). Each cell reports the mean over three runs, with the sample SD in parentheses below.

Wave	Lasso MSE	ENet MSE	$\Delta$	Winner
W26	.0098 $\pm$ .0002	.0098 $\pm$ .0002	.0000	Tie
W27	.0218 $\pm$ .0010	.0238 $\pm$ .0027	.0020	Lasso
W29	.0156 $\pm$ .0049	.0156 $\pm$ .0049	.0000	Tie
W32	.0160 $\pm$ .0019	.0160 $\pm$ .0019	.0000	Tie
W34	.0111 $\pm$ .0011	.0111 $\pm$ .0011	.0000	Tie
W36	.0191 $\pm$ .0042	.0187 $\pm$ .0037	.0004	ENet

Appendix

Table 4: Comparison of lasso and elastic net regression for P2P across 14 ATP waves. Values are mean

\pm

sample SD over three runs. Both methods yield nearly identical performance, demonstrating robustness to regularization choice.

Metric	Vanilla	Vanilla	PERSONA	PERSONA
		(1000 $\to$ 300)		(1000 $\to$ 300)
Avg. Entropy $\uparrow$	0.3172 $\pm$ 0.0069	0.3161	0.2871 $\pm$ 0.00026	0.2844
Test MSE (Lasso) $\downarrow$	0.0285 $\pm$ 0.00490	0.0266	0.0362 $\pm$ 0.00116	0.0395
Test MSE (Enet) $\downarrow$	0.0286 $\pm$ 0.00476	0.0266	0.0339 $\pm$ 0.00124	0.0387
Cost (USD) $\downarrow$	0.8649 $\pm$ 0.0070	0.8833 †	1.7129 $\pm$ 0.000016 *	1.7203 *†

Appendix

Table 5: Comparison of Vanilla, PERSONA and their clustering-based variants where 1000 agents are reduced to 300 via

k

-means clustering.

Metric	WVS W7 US	WVS W7 GB	WVS W7 HK
Avg. Entropy $\uparrow$	0.5720 $\pm$ 0.0149	0.5710 $\pm$ 0.0044	0.5478 $\pm$ 0.0211
Cost (USD) $\downarrow$	1.7551 $\pm$ 0.0097	1.7095 $\pm$ 0.0031	1.7365 $\pm$ 0.0080
Test MSE (Lasso) $\downarrow$	0.0149 $\pm$ 0.0005	0.0141 $\pm$ 0.0014	0.0242 $\pm$ 0.0013
Test 1-JSD (Lasso) $\uparrow$	0.8021 $\pm$ 0.0042	0.7892 $\pm$ 0.0052	0.7532 $\pm$ 0.0056
Test MCPA (Lasso) $\uparrow$	0.6667 $\pm$ 0.0152	0.6481 $\pm$ 0.0321	0.5789 $\pm$ 0.0263
Test EMD (Lasso) $\downarrow$	0.1073 $\pm$ 0.0032	0.1032 $\pm$ 0.0072	0.1423 $\pm$ 0.0035

Appendix

Table 6: P2P performance on WVS Wave 7 (US, Great Britain, Hong Kong SAR) under a 7:1.5:1.5 question split, using gemini-2.0-flash as the backend. Reported are mean

\pm

sample SD over three repeated runs.

Metric	ATP (14 waves)	WVS US
Test MSE (Lasso) $\downarrow$	0.014 $\pm$ 0.004	0.015 $\pm$ 0.001
Avg. Entropy $\uparrow$	0.53 $\pm$ 0.06	0.57 $\pm$ 0.01
Cost (USD) $\downarrow$	0.8 $\pm$ 0.2	1.76 $\pm$ 0.01

Appendix

Table 7: Comparison of P2P performance on ATP (US) and WVS (US). ATP values are mean

\pm

sample SD over 14 waves; WVS values are mean

\pm

sample SD over 3 runs.

Metric	Avg. Entropy $\uparrow$	Test MSE $\downarrow$	Test MSE $\downarrow$	Cost (USD) $\downarrow$
		(Lasso)	(Enet)
Original	0.4451 $\pm$ .0256	0.0087 $\pm$ .0026	0.0082 $\pm$ .0025	0.9708 $\pm$ .0036
Survey patch off	0.4378 $\pm$ .0134	0.0104 $\pm$ .0024	0.0102 $\pm$ .0017	0.9785 $\pm$ .0122
Question patch off	0.3989 $\pm$ .0203	0.0114 $\pm$ .0016	0.0109 $\pm$ .0015	0.9698 $\pm$ .0011
Freeform off	0.3953 $\pm$ .0097	0.0128 $\pm$ .0008	0.0123 $\pm$ .0012	0.9688 $\pm$ .0015
max attributes = 20	0.4363 $\pm$ .0061	0.0124 $\pm$ .0046	0.0115 $\pm$ .0041	0.9880 $\pm$ .0167

Appendix

Table 8: Attribute ablation. Reported are mean

\pm

sample SD over three repeated runs.

Table 9: Performance of P2P with varying endowment budget. The average question entropy, test MSE (Lasso and ElasticNet), and generation cost are reported. All values are mean

\pm

sample SD over three repeated runs.

Model	Avg. Entropy	Test MSE (Lasso)	Test MSE (Enet)	Cost (USD)
GPT-4.1-mini	0.4708	0.0090	0.0088	3.8342
	(0.0168)	(0.0014)	(0.0014)	(0.0449)
GPT-4.1-nano	0.3996	0.0226	0.0213	0.9265
	(0.0154)	(0.0042)	(0.0048)	(0.0043)
Qwen	0.1932	0.0209	0.0201	–
	(0.0403)	(0.0035)	(0.0029)

Appendix

Table 10: Comparison of P2P performance on W42 using different model backends. Reported are average question entropy, test MSE (Lasso and ElasticNet), and generation cost. For each model, we fix the endowment budget at 300. Values are means over three repeated runs, with sample SDs in parentheses.

Metric	Without regression	With regression
Test MSE $\downarrow$	0.0332 $\pm$ 0.0024	0.0087 $\pm$ 0.0026

Appendix

Table 11: Comparison of P2P performance with and without the regression-based aggregation stage on ATP W42. In the ablation, agent responses from active endowment generation are combined by simple averaging, while the regression setting uses lasso. Reported are mean

\pm

sample SD over three repeated runs.

Weighting	Train MSE $\downarrow$	Test MSE $\downarrow$	Train MSE $\downarrow$	Test MSE $\downarrow$
Scheme	(Lasso)	(Lasso)	(Enet)	(Enet)
Default	0.00497 $\pm$ .00156	0.00870 $\pm$ .00256	0.00403 $\pm$ .00144	0.00821 $\pm$ .00249
Block	0.00475 $\pm$ .00232	0.00868 $\pm$ .00257	0.00442 $\pm$ .00236	0.00806 $\pm$ .00278

Appendix

Table 12: Sensitivity check of weighting schemes. Reported are mean

\pm

sample SD over 3 runs.

With mass penalty
Configuration	MSE $\downarrow$	1-JSD $\uparrow$	EMD $\downarrow$	MCPA $\uparrow$	Agents
No uniform component (default)	0.0142	0.8300	0.0657	0.6764	105.7
+ Reference respondent	0.0130	0.8365	0.0699	0.6790	51.9
+ Intercept	0.0128	0.8403	0.0679	0.6928	60.4
Without mass penalty
No uniform component	0.0146	0.8225	0.0672	0.6749	43.7

Appendix

Table 13: Stage 2 configuration ablation on the main panel (14 ATP waves, three repeated runs each; lasso; test split; Gemini-2.0-flash). Reported are means over the 42 runs. Agents is the mean number of agents with non-zero weight.

Setup	Train+Val Entries	Test Entries	Model backend
	(Q1 + Q2)	(Q3)
P2P (US)	185 (150 + 35)	59	gemini-2.0-flash
P2P (GB)	172 (140 + 32)	60	gemini-2.0-flash
P2P (HK)	185 (150 + 35)	59	gemini-2.0-flash
Cao et al. (C1: 46 countries)	8427 (6841 + 1586)	2719	Llama3-8B-Instruct

Appendix

Table 14: Different training regimes for P2P and [ cao-etal-2025-specializing ] .

Setup	Method	Test MSE $\downarrow$	Test 1-JSD $\uparrow$	Test MCPA $\uparrow$	Test EMD $\downarrow$
P2P (US)	Lasso	0.0217 $\pm$ .0014	0.7473 $\pm$ .0031	0.6271 $\pm$ .0169	0.1174 $\pm$ .0051
P2P (US)	Enet	0.0218 $\pm$ .0019	0.7488 $\pm$ .0044	0.6384 $\pm$ .0098	0.1177 $\pm$ .0072
P2P (GB)	Lasso	0.0249 $\pm$ .0009	0.7416 $\pm$ .0055	0.4889 $\pm$ .0347	0.1296 $\pm$ .0045
P2P (GB)	Enet	0.0249 $\pm$ .0009	0.7463 $\pm$ .0067	0.5111 $\pm$ .0096	0.1294 $\pm$ .0048
P2P (HK)	Lasso	0.0284 $\pm$ .0023	0.7115 $\pm$ .0125	0.4972 $\pm$ .0353	0.1507 $\pm$ .0122
P2P (HK)	Enet	0.0290 $\pm$ .0023	0.7132 $\pm$ .0150	0.4859 $\pm$ .0427	0.1528 $\pm$ .0105

Appendix

Table 15: P2P vs. cao-etal-2025-specializing on WVS (Q3). Reported are mean

\pm

sample SD over 3 repeated runs.

P2P , gemini-2.0-flash panel (Table 1 bases)
Configuration	WD $\downarrow$	MSE $\downarrow$	1-JSD $\uparrow$	MCPA $\uparrow$
Mass penalty (default)	0.1231 $\pm$ .0014	0.0253 $\pm$ .0007	0.8257 $\pm$ .0028	0.6714 $\pm$ .0010
+ Reference respondent	0.1127 $\pm$ .0013	0.0192 $\pm$ .0004	0.8504 $\pm$ .0016	0.6817 $\pm$ .0070
+ Intercept	0.1155 $\pm$ .0013	0.0204 $\pm$ .0003	0.8477 $\pm$ .0013	0.6690 $\pm$ .0024
No mass penalty + intercept	0.1148 $\pm$ .0009	0.0187 $\pm$ .0003	0.8500 $\pm$ .0010	0.6897 $\pm$ .0054
P2P , GPT-4o/GPT-4o-mini panel

Appendix

Table 16: Subgroup prediction on the OpinionQA benchmark of suh-etal-2025-language (13 ATP waves, 22 subgroups, 208 test questions, 4,572 question–subgroup cells). P2P reports mean

\pm

sample SD over three agent bases (lasso); SubPOP-FT uses the released adaptor. All four metrics, including MSE, are computed on P2P ’s clipped and renormalized predictions. WD is the normalized ordinal Wasserstein distance used by suh-etal-2025-language .

Configuration	W6 subgroup refit	W7 subgroup fit $\to$ W6	W7 population fit $\to$ W6
Mass penalty (default)	0.0127 $\pm$ .0008	0.0129 $\pm$ .0001	0.0123 $\pm$ .0008
+ Reference respondent	0.0088 $\pm$ .0010	0.0090 $\pm$ .0002	0.0092 $\pm$ .0003
+ Intercept	0.0108 $\pm$ .0006	0.0112 $\pm$ .0001	0.0116 $\pm$ .0007
No mass penalty + intercept	0.0093 $\pm$ .0010	0.0096 $\pm$ .0002	0.0101 $\pm$ .0001
SubPOP-FT (zero-shot)	0.0249

Appendix

Table 17: Subgroup prediction on WVS (US; 16 held-out W6 questions, 10 subgroups common to W6 and W7). Reported is test MSE, mean

\pm

sample SD over the three gemini-2.0-flash bases of Table 6 (lasso). SubPOP-FT is applied zero-shot.

	Calibration anchors
Configuration	5	10	20	40	60	72
Test questions (16)
Persistence (W7 marginals)	3.64
Direct refit, with mass penalty
No uniform component (default)	11.64	11.56	11.28	10.93	11.19	11.01
	(0.71)	(0.72)	(0.92)	(0.80)	(0.65)	(0.42)

Appendix

Table 18: W6 reconstruction via the W7 proxy basis (WVS, US; three bases

\times

ten random splits). Reported is mean MSE (in units of

10^{-3}

) over the 30 basis–split cells, with SD in parentheses (for persistence, across the ten splits). Direct refits predict the W6 marginals with elastic net. Drift refits predict the W6

-

W7 change using W7 as the reference, with agent weights summing to zero. All fits use option-level weighting, with penalties chosen on 19 fixed validation questions. Persistence uses the W7 marginals unchanged. Bold indicates the lowest MSE per column and panel.

Method	Representation of heterogeneity	Supervision required	What is adapted
GPO [ zhao2024group ]	In-context group preference examples	Cross-group preference data (meta-training) and target-group context examples	Preference module (meta-trained offline)
Modular Pluralism [ feng-etal-2024-modular ]	Specialized community models	Community-specific corpora	Community LMs (offline)
SPICA [ chen2024spica ]	Retrieved group perspectives/scenarios	Group-labelled scenario resources	None (in-context)
PERSONA [ castricato_2025_persona ]	Synthetic demographic personas with LLM-enriched features	Census-derived demographic information	None (in-context)
cao-etal-2025-specializing	Locale-conditioned predictor	${\sim}$ 8.4K pooled response entries from 46 countries	Fine-tuned model weights
SubPOP-FT [ suh-etal-2025-language ]	Subgroup-conditioned predictor	70K subgroup–response pairs from 53 ATP waves	LoRA weights

Appendix

Table 19: How P2P differs from related methods. Existing methods take heterogeneity from external representations or learn it in model parameters; P2P actively constructs it from the task at hand.

	SubPOP-FT	Cao et al.	P2P
Target populations	22 U.S. demographic subgroups	46 country populations	Defined by the given survey
Applying to a new population	New training corpus and fine-tuning	Locale supplied as model input	Re-run the pipeline (new survey) or refit Stage 2 weights (same survey); locale conditioning not required (Section 4.4 )
Adaptation	LoRA fine-tuning; GPU	Fine-tuning; GPU	Inference-time regression; CPU
What varies, training to test	Fielding period, respondents, topic	Topic only	Questions only
Temporal drift	Not modelled	Not modelled	Weights refit per survey occasion

Appendix

Table 20: Comparison of suh-etal-2025-language (SubPOP-FT), cao-etal-2025-specializing , and P2P .

Preference Modeling

See all 45 papers

Sep 28, 2026

Simulating Respondents, Not Single Questions: Coherent Survey Generation with Large Language Models

Large language models are increasingly used to simulate response distributions in social surveys. Prior work has achieved accurate population-level simulation for individual questions. Real questionnaires, however, ask each respondent a sequence of related questions. A simulated respondent should show coherent preferences across the whole questionnaire, not merely accurate distributions for isolated items. Existing single-item methods cannot accurately reproduce how the same person answers a complete survey. We propose FullRespondent-LLM (FR-LLM), which fine-tunes two specialized LLMs: a marginal model for each item's response distribution and a respondent-level autoregressive model for dependencies across answers. Marginal-Constrained Joint Projection (MCJP) then projects the autoregressive joint distribution onto the set satisfying the item-level marginals learned by the first model. This yields complete questionnaires with realistic cross-item relationships while retaining strong item-level accuracy. On two real-world social survey datasets, FR-LLM more accurately reproduces multi-question response patterns, maintains competitive single-item accuracy, and generalizes better to unseen populations and questions. In a small commercial-survey dataset, we use simulated responses to make pricing and stocking decisions; FR-LLM achieves the highest realized profit.

Ji Huang, Mengfei Li, Shuai Shao

School of Computer Science, University of Science and Technology of China · School of Management, Fudan University

Large Language Model-Based Social Simulation Conditional Generative Modeling

Ask PDF

Oct 5, 2026

Synthetic Cultural Agents from Aggregate Anchors

Population prompts are widely used to generate synthetic survey responses, but they combine information supplied at inference with associations already encoded during pretraining. We introduce an alternative construction that maps declared aggregate preference anchors into group-indexed choice policies. For each population, the signs of six Global Preferences Survey (GPS) coordinates deterministically label a shared bank of paired synthetic responses, and Direct Preference Optimization fits a parameter-efficient adapter to those comparisons. We evaluate the adapters on candidate World Values Survey (WVS) items using prompts that omit country names and distinguish four questions: recovery of the imposed labels, transfer of the anchor signal to new text, coherence between the GPS anchors and human WVS responses, and agreement between adapter and human scores. The adapters recover the imposed pairwise labels. On a purposively selected sixteen-country development panel, adapter trust scores completely separate the two GPS-sign groups and have a rank correlation of (0.74) with continuous GPS trust scores. Human-GPS and adapter-human associations remain unresolved on the same panel, and results for the other preference dimensions are heterogeneous. These findings show that an anchored policy can retain a declared aggregate signal without thereby reproducing human response patterns. The contribution is therefore both an inspectable construction and an evaluation framework that separates anchor transfer from human criterion agreement.

Augusto Gonzalez-Bonorino, Kseniia Biriukova, Monica Capra

Department of Economics, Arizona State University · EconLLM Lab · Department of Information Systems, Arizona State University +1

Large Language Model-Based Social Simulation Synthetic Data Generation for Language Models

Ask PDF

May 15, 2026

Improving Cross-Cultural Survey Simulation with Calibrated Value Personas

Large language models (LLMs) are increasingly used to simulate human opinions and survey responses, but their ability to reproduce population responses across cultures remains limited. Existing persona-based prompting methods typically rely on sociodemographic or personality traits, which are only indirect proxies for the values that shape human responses. We propose a value-based persona construction method that derives textual descriptors from survey responses capturing core cultural dimensions. By sampling value profiles from target populations and aggregating LLM responses across personas, we obtain population-level predictions grounded in observed value distributions. We further introduce a calibration procedure that improves response diversity while preserving estimated opinions. We show that our approach reduces prediction error across countries, with the largest improvements observed in underrepresented populations. This substantially narrows the performance gap between countries aligned with dominant LLM priors and those that are less represented in training data, while also yielding response distributions that closely match human diversity.

Axel Abels, Elias Fernandez Domingos, Apurva Shah +1

1Machine Learning Group, Universit´e Libre de Bruxelles · 2AI Lab, Vrije Universiteit Brussel · 3FARI Institute, Universit´e Libre de Bruxelles - Vrije Universiteit Brussel +2

Language Model Calibration Cultural Bias in Language Models Persona Prompting Prompt Engineering

Ask PDF

Abstract

Preference Modeling

Explore similar work

Simulating Respondents, Not Single Questions: Coherent Survey Generation with Large Language Models

Synthetic Cultural Agents from Aggregate Anchors

Improving Cross-Cultural Survey Simulation with Calibrated Value Personas