Paper deep dive

Data-Driven Personas for Survey Simulation: Insights into Simulation Alignment Across Data-Access Regimes

Dongryeol Lee, Weronika Łajewska, Leonardo Perelli, Saab Mansour

Year: 2026Venue: arXiv preprintArea: cs.AIType: PreprintEmbeddings: 88

Intelligence

Status: not_run | Model: - | Prompt: - | Confidence: 0%

Entities (0)

No extracted entities yet.

Relation Signals (0)

No relation signals yet.

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:However, many existing steering approaches rely on target-domain human data for fine-tuning or prompting that is costly to collect and raises privacy concerns. In this paper, we study demographic group-level survey simulation, where personas induced from heterogeneous, anonymized public behavioral data condition agents that simulate responses of individuals from specific demographic groups. We examine whether representative personas can be induced from diverse sources and analyze how the domain, scale, and granularity of the source data affect survey simulation alignment. We find that personas induced from out-of-domain sources rarely outperform simulations conditioned only on basic demographic information, largely due to population mismatch. However, when personas are accurately assigned to the target demographic groups, alignment improves substantially. Finally, personas induced from target-domain survey data generalize better as more survey question history becomes available, suggesting that richer behavioral evidence enables more stable persona trait inference that transfers to better unseen questions simulation alignment.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

85,351 characters extracted from source content.

Expand or collapse full text

Data-Driven Personas for Survey Simulation: Insights into Simulation Alignment Across Data-Access Regimes Dongryeol Lee ∗ Seoul National University drl123@snu.ac.kr Weronika Łajewska Amazon lajewska@amazon.com Leonardo Perelli Amazon lperelli@amazon.com Saab Mansour Amazon saabm@amazon.com Abstract Large language models (LLMs) offer new op- portunities for public opinion research by en- abling early prediction of survey responses, po- tentially reducing the cost and time of tradi- tional surveys. However, many existing steer- ing approaches rely on target-domain human data for fine-tuning or prompting that is costly to collect and raises privacy concerns. In this paper, we study demographic group-level sur- vey simulation, where personas induced from heterogeneous, anonymized public behavioral data condition agents that simulate responses of individuals from specific demographic groups. We examine whether representative personas can be induced from diverse sources and ana- lyze how the domain, scale, and granularity of the source data affect survey simulation align- ment. We find that personas induced from out- of-domain sources rarely outperform simula- tions conditioned only on basic demographic in- formation, largely due to population mismatch. However, when personas are accurately as- signed to the target demographic groups, align- ment improves substantially. Finally, personas induced from target-domain survey data gen- eralize better as more survey question history becomes available, suggesting that richer be- havioral evidence enables more stable persona trait inference that transfers to better unseen questions simulation alignment. 1 Introduction Surveys are an essential tool for understanding pub- lic opinion and informing decisions in public policy, industry, and social science research. Because at- titudes and preferences evolve over time and vary across demographic groups, organizations often rely on repeated and large-scale surveys to capture these differences. However, collecting survey data ∗ Work done as an intern at Amazon. Inferred Demographic Attributes Persona You are a committed progressive who identifies as liberal and strongly aligns with the Democratic Party... Political Party: Democrat / highconfidence Region: South/lowconfidence c Group㎿level Testset Evaluation Human Responses Agent Prediction Human Behavior Data Persona Generation •Survey: OpinionQA •Survey: GSS •Survey: SocSci210 •Movie Review •E㎿commerce Question:Do you favor or oppose the death penalty for persons convicted of murder? Options: (A): Favor (B): Oppose Subgroup: ㏛Political Party: Democrat㏜ (A) (B) (A) (B) Figure 1: Data driven persona induction and survey sim- ulation: In our framework, we generate personas from behavioral data across different domains and predict responses for individuals in target demographic groups. is costly, time-consuming, and logistically demand- ing. Recent advances in large language models (LLMs) have enabled their use for survey simu- lation, where models are prompted to predict the opinions of different demographic groups (Wang et al., 2025; Gao et al., 2024; Manning et al., 2024; Bail, 2024). This approach offers a scalable alterna- tive to traditional user studies while enabling faster exploration of public opinion dynamics (Argyle et al., 2023; Aher et al., 2023; Hewitt et al., 2024). Prior work on survey simulation has primar- ily explored prompt engineering (Santurkar et al., 2023; Kim and Lee, 2023), persona-based condi- tioning (Park et al., 2024a), and fine-tuning (Cao et al., 2025; Kolluri et al., 2025; Suh et al., 2025) to improve the alignment between simulated and human responses. However, much of this work re- lies on target-domain human data for learning and control—for example, by using interviews or other arXiv:2610.05828v1 [cs.AI] 5 Oct 2026 individual-level records in prompts (Park et al., 2024a), or fine-tuning models on data drawn from the same source as the evaluation set (Cao et al., 2025; Kolluri et al., 2025; Suh et al., 2025). Col- lecting such data is resource-intensive and often constrained by privacy regulations, consent require- ments, and governance frameworks, while also pro- viding limited coverage of underrepresented sub- populations. In this work, we investigate the impact of synthetic personas induced from heterogeneous sources of human behavioral data on steering LLM- based survey simulations. 1 Our goal is to un- derstand when such personas, defined as natural- language profiles distilled from behavioral traces, improve the alignment of survey simulation. This question is important in practice because practi- tioners often have uneven access to persona source data. In some settings, only external behavioral data are accessible (Low access / out-of-domain); in others, additional information such as ground- truth (GT) demographics of participants may be accessible (Mid access / out-of-domain with GT demographics); and in the most favorable cases, part of the target survey data itself may be available (High access / in-domain). To reflect these realis- tic constraints, we frame persona-conditioned sur- vey simulation in terms of data-access regimes and study how the effectiveness of personas changes as stronger forms of supervision become available. Accordingly, we treat persona induction from behavioral data and downstream survey simulation as separate tasks. We first induce personas from dif- ferent sources of behavioral data, then use them to condition individual agents, and evaluate the result- ing aggregated predictions against human response distributions at the demographic-group level (e.g., the Democrat group distribution in Figure 1). 2 Across survey benchmarks, we find that syn- thetic personas are not uniformly beneficial for simulation—their effectiveness depends strongly on the available data and the degree of source– target alignment.Nevertheless, the value of persona-conditioned simulation should not be 1 All experiments use publicly available datasets with anonymized behavioral traces, and no personally identifiable information was used or collected in this research. To support reproducibility, we release the code athttps://github.com/ amazon-science/data-driven-personas , but not raw user-level traces, record-linked personas, or intermediate arti- facts that could increase re-identification risk. 2 We do not aim to replicate individual behavior, but to statistically approximate group-level responses. judged solely by aggregate alignment. Persona- based simulation can also offer an important advan- tage in interpretability: each simulated response is linked to a specific persona induced from be- havioral evidence, and aggregating over multiple personas allows the simulator to represent within- group heterogeneity rather than collapsing a group into a single demographic descriptor. We there- fore view personas not only as a potential means of improving aggregate alignment, but also as a way to make survey simulation more behaviorally grounded, more interpretable, and more sensitive to within-group variation. More broadly, our results provide practical guidance for deploying survey simulation under realistic data constraints, show- ing that effective simulation depends not only on the richness of persona information, but also on the type of behavioral data, accuracy of persona to group assignment, and source–target population mismatch (i.e., the degree to which the source be- havioral population resembles the target survey re- spondents). Our contributions are summarized as follows: •We show that personas induced from out-of- domain data do not generalize well to survey sim- ulation, largely because demographic attribute misassignment causes personas to be mapped to inappropriate target groups. • We show that personas induced from in-domain data generalize better to held-out questions as more history is provided. They also yield the strongest and most stable alignment across dif- ferent sizes of demographic groups, with espe- cially large gains for smaller, underrepresented groups, and perform best across all demographic attributes. 2 Related Work Survey simulation.Recent work explores LLMs as a lower-cost alternative to large-scale human data collection, demonstrating their ability to re- produce some group-level response patterns in sur- veys and behavioral, economic, and political set- tings (Argyle et al., 2023; Aher et al., 2023; Hor- ton, 2023; Hewitt et al., 2024). However, subse- quent studies raised concerns about validity, ro- bustness, and representational bias (Dominguez- Olmedo et al., 2024; Bisbee et al., 2024). Subse- quent efforts proposed prompting and elicitation methods, including sociodemographic prompting, persona-based conditioning, and elicitation of re- sponse distributions (Santurkar et al., 2023; Moon et al., 2024; Park et al., 2024a; Manning et al., 2024; Kwok et al., 2024; Sun et al., 2024). Yet capturing within-group heterogeneity and avoiding stereo- typical predictions remain open challenges (Park et al., 2024b; Kapania et al., 2025; Cheng et al., 2023), and fine-tuning approaches rely on survey or experimental response data for adaptation (Cao et al., 2025; Kolluri et al., 2025; Suh et al., 2025). In contrast, we focus on simulation of group-level response distributions and examine how condition- ing signals transfer across domains and data-access regimes. Synthetic personas. Synthetic persona genera- tion constructs artificial user profiles to support downstream tasks such as political and economic simulation (Tseng et al., 2024; Chen et al., 2024; Horton, 2023). Earlier approaches often relied on survey or census data (Argyle et al., 2023; Chang et al., 2025), demographic profiles (Hwang et al., 2023; Santurkar et al., 2023), or political identities (Simmons, 2023) to steer model behav- ior. With LLMs, recent work has moved toward descriptive personas derived from richer sources, including web text (Ge et al., 2024), interview transcripts (Park et al., 2024a), personal narra- tives (Moon et al., 2024), and shopping histo- ries (Mansour et al., 2025), with behavioral-data- grounded personas also supporting A/B test simula- tion (Benomar et al., 2026). Other work further ex- plores automatically refined personalization: few- shot personalization (Kim and Yang, 2025) and col- laborative data refinement (Sun et al., 2025). These methods provide richer conditioning signals, but approaches grounded in individual records may de- pend on costly or sensitive target-domain data, and demographic prompting can yield caricatured sim- ulations (Cheng et al., 2023). In contrast, we focus on data-driven personas induced from external be- havioral sources to steer distribution-level survey simulation, enabling out-of-domain deployment without individual-level target-survey records. 3 Survey Simulation under Different Data-access Regimes In this work, we study agentic survey simulation, in which a pool of LLM-based agents predict how humans would respond to survey questions (Ar- gyle et al., 2023; Wang et al., 2025). Our central question is whether—and under what conditions— Input Behavioral Data for each individual GSS / OpinionQA ㏝Q1㏞: Do you favor or oppose the death penalty for persons convicted of murder? Favor; Oppose ㏝R1㏞: Favor ㏝Q2㏞: Do you think the use of marijuana should be made legal or not? Should; Should not [R2]: Should Movie Title 1: DarStar: HR Gigers Welt ㏙2014㏚ Review 1:An Awful Documentary Instead of being a joyous celebrationof the life... Title 2: Tales of Manhattan ㏙1942㏚ Review 2: Astaire 㖈 Powell Dance Up A Story... GT Demographic Attributes ㏙Optional㏚㏛REGION:South㏜, ㏛AGE: 30㎿49㏜㏛Gender: Male㏜, ㏛Education: Postgraduate㏜ N/A Persona Generation Output ㏙i㏚ Persona You are a committed progressive who identifies as extremely liberal and strongly aligns with the Democratic Party. You... You are a thoughtful and empathetic person with a strong sense of justice and fairness. You\㑄re particularly attuned to power... ㏙i㏚ Inferred Demographic Attributes ㏙㏛Key: Value / confidence level㏜㏚㏛REGION: Midwest / low㏜㏛AGE: 30㎿49 / high㏜㏛GENDER: Female / med㏜ ... ㏛REGION: Midwest / low㏜㏛AGE: 30㎿49 / low㏜㏛GENDER: Female / med㏜ ... Figure 2: Persona generation examples across domains. synthetic personas, defined as natural-language pro- files inferred from behavioral data, improve simula- tion alignment for survey responses. In our setting, agents are conditioned on synthetic personas, while evaluation is performed on aggregated response dis- tributions for each demographic group (Suh et al., 2025; Cao et al., 2025; Kolluri et al., 2025). 3.1 Persona-conditioned Simulation LetQbe a set of survey questions. Each question q ∈ Qhas answer optionsX q = x 1 , . . . , x N q . For a demographic groupg, defined by a single attribute–value pair (e.g.,AGE=30–49), empirical human responses induce a subgroup distribution U q,g overX q . Our goal is to produce a simulated distribution S q,g that is closely aligned with U q,g . To improveS q,g , we condition each agent on a synthetic persona. Our framework proceeds in two stages: 1) persona generation and 2) persona- conditioned survey simulation. In the first stage, we induce a pool of personas from behavioral data. Each persona is a natural-language description of the individual’s characteristics and is associated with inferred demographic attributes accompanied by corresponding confidence scores (see Figure 2), enabling it to be matched to a target demographic groupg. In the second stage, these induced per- sonas condition individual agents to simulate re- sponses for a target survey, and the resulting predic- tions are aggregated for each demographic group. For a target questionqand groupg, we select per- sonas that match the demographic attribute defining g, promptN A LLM agents—one per persona—to produce a verbalized distribution (Hu et al., 2026) over the answer options, and then aggregate out- puts: 3 S q,g (x) = 1 N A P N A j=1 S (j) q,g (x), x ∈ X q . This formulation separates persona construction from downstream evaluation, allowing the same simulation protocol to be instantiated with differ- ent persona sources and different levels of access to target-domain information. 3.2 Data Access Regimes We organize our experiments based on the infor- mation available for persona construction and for assigning personas to demographic groups. Each data-access regime motivates a corresponding re- search question. Low-access / out-of-domain.This regime repre- sents situations where practitioners have no access to target-survey data. Personas must instead be in- duced from external sources, such as non-survey behavioral data, survey data from other datasets or waves, or the LLM’s parametric knowledge. It mimics scenarios where only general behavioral evidence is available, testing whether such out-of- domain personas can still produce meaningful pre- dictions for a target survey. This motivates RQ1: Can personas induced from out-of-domain behav- ioral data improve survey simulation alignment? Mid-access / out-of-domain setting with GT de- mographics This regime reflects settings where practitioners have behavioral data from external sources along with GT demographic information. Personas can be induced using both behavioral traces and demographic attributes. This scenario captures cases where demographic alignment is fea- sible even without direct access to the target survey, allowing us to study how correct persona–group as- signment affects simulation alignment. This moti- vates RQ2: How does access to ground-truth demo- graphic information in out-of-domain behavioral data improve survey simulation alignment? High-access / in-domain.This regime represents scenarios where practitioners have partial access to the target survey itself. Personas are induced from these in-domain responses and used to predict held-out questions within the same survey. It cap- tures the potential of leveraging partial in-domain data to improve simulation for unseen questions, 3 Based on the findings of our pilot study (Appendix A), we setN A = 30, select the highest-confidence matching personas for each demographic group, and use verbalized distributions to represent agent outputs. motivating RQ3: Can in-domain personas induced from partial survey responses generalize to predict held-out questions? 4 Experimental Setup We now instantiate the study formulation described in Section 3 with concrete target datasets, persona sources, baselines, and evaluation metrics. 4.1 Target Survey Datasets Construction We evaluate survey simulation on five publicly available survey datasets or waves: OpinionQA Waves 26 and 29 (Santurkar et al., 2023), the General Social Survey (GSS) (Davern et al., 2024), and two SocSci210 subsets, MP3DR and NJ5DX (Kolluri et al., 2025). 4 Each dataset pro- vides respondent-level demographic attributes and individual survey responses, enabling construction of group-level empirical response distributions. Following Section 3, we construct demographic group-level test instances by aggregating responses from respondents who share a given demographic attribute value. This yields test instances of the form (q i ,x k , g, U i,g ). 5 4.2 Persona Sources We construct personas using three types of data. Model knowledge.Given a target attribute defin- ing the demographic groupg, we prompt Claude 3.7 Sonnet (Anthropic, 2025) to generate a diverse set of personas consistent with that group. This provides a source-free reference point that relies only on the model’s parametric knowledge. Non-survey behavioral data. We induce per- sonas from anonymized movie review histo- ries (Yoon et al., 2024) and e-commerce shopping histories (Berke et al., 2024). This instantiates the low-access setting where only external out-of- domain behavioral traces are available. Survey behavioral data. Personas are gener- ated from anonymized respondent-level survey re- sponses. When the source survey differs from the target, this represents an out-of-domain setting. When the source survey is the same as the target sur- vey, we split the questions into persona-induction 4 OpinionQA waves and SocSci210 subsets differ in topics and populations and are therefore treated as separate surveys. 5 Appendix D covers statistics of both target dataset and persona-source datasets and shows example of test instance. and evaluation subsets, using the former as the be- havioral data source and the latter as held-out test questions, representing the in-domain setting. 4.3 Persona Generation and Selection For data-driven persona generation, we use Claude 3.7 Sonnet (Anthropic, 2025) as the persona gener- ator. As shown in Figure 2, given an individual’s behavioral record, the model produces: (i) a natural- language persona for simulation, and (i) inferred demographic attributes for 11 keys (CREGION, AGE, SEX, EDUCATION, MARITAL, RELIG, RELIGATTEND, POLPARTY, POLIDEOLOGY, INCOME, and RACE) with confidence scores. Each key has a fixed closed set of options (e.g., AGE: 18–29 | 30–49 | 50–64 | 65+; POLPARTY: Democrat | Republican | Independent | Other); full option sets are listed in Appendix B. For each demographic groupg, we assign per- sonas using their inferred demographic attributes, then rank candidates by confidence score and select the topN A = 30per group for simulation. This confidence-based selection is motivated by our pi- lot study (Appendix A), which shows that high- confidence extraction yields substantially higher demographic accuracy (e.g., 0.81 vs. 0.39 on GSS; Table 7), and that simulation performance largely saturates beyond 30 personas. For settings where ground-truth demographic labels are available for source individuals (Mid ac- cess / out-of-domain with GT demographics), we consider two additional variants. In GT assign- ment, ground-truth demographics are used only for the assignment of personas to different demo- graphic groups. In GT generation, ground-truth demographic information is also provided as an additional data source during persona induction. 4.4 Baselines and Simulation Protocol We compare the persona-conditioned simulation against three baselines following SimBench (Hu et al., 2026). 6 B1 (No demographic context). Each agent re- ceives only the question and answer options. B2 (Country-level context).Each agent receives a short demographic context (e.g. You are from USA) in addition to the question. 6 All baselines use the same individual-agent simulation and group-level aggregation procedure. B3 (Single-attribute context). Each agent re- ceives an explicit value of a demographic attribute (e.g. You are Male) along with the question. Simulation protocol.For each target questionq and demographic groupg, we promptN A = 30 LLM agents in parallel, each conditioned on a dis- tinct persona selected as described in Section 4.3. Each agent produces a verbalized probability dis- tribution over the answer options (e.g.,A: 30%, B: 40%, C: 30%), and the resulting distributions are aggregated via simple averaging as described in Section 3. This aggregation approximates within- group heterogeneity by pooling over a diverse set of persona-conditioned agents rather than collapsing each group to a single representative response. 4.5 Evaluation Metrics We quantify the misalignment between the hu- man demographic group-level response distribu- tionU i,g (x)and the simulated response distribu- tionS i,g (x)using two complementary metrics: To- tal Variation Distance (TVD) (Hu et al., 2026; Ła- jewska et al., 2026) and 1-Wasserstein distance (WD) (Suh et al., 2025; Moon et al., 2024) (see Appendix C for details). For each test instance (q i , g), we compute the distance betweenU i,g and S i,g and report the average across all questions and demographic groups. Lower values indicate closer alignment for both metrics. 5 Results We evaluate persona-conditioned simulation across five target survey datasets. 7 For each persona source dataset, we first construct a persona pool, then condition simulation on individual personas, and evaluate performance at the demographic group level. 8 5.1 Low-access Regime / Out-of-domain Setting We begin with a practically motivated low-access setting in which only human behavior traces from non-target sources are available (RQ1). We con- sider persona sources including external behavioral corpora (movie reviews and shopping histories) and survey respondent data from other datasets/waves. 7 The simulation is executed with Claude 3.7 Sonnet using N A = 30agents for each demographic group. We provide additional results with open-weight LLMs in Appendix G. 8 Note that the same persona pool can be reused for differ- ent target datasets in the out-of-domain setting. Baseline/Persona Source GSSOpinionQA W26 OpinionQA W29 SocSci210 (MP3DR) SocSci210 (NJ5DX) TVDWDTVDWDTVDWDTVDWDTVDWD B10.169 0.322 0.1890.2990.1710.2620.3110.4740.2410.678 B20.1600.3010.1730.2740.1390.2040.3140.5520.2040.560 B30.151 0.280 0.178 0.2870.1300.1940.3230.5830.2040.538 Model Knowledge0.1600.303 0.1900.3110.1570.2370.3300.5790.1910.443 Movie Reviews0.169 0.326 0.1970.3180.1760.2630.3150.5920.2030.506 Shopping History0.163 0.312 0.1890.3010.1600.2400.3090.5320.2080.530 SocSci210 MP3DR0.186 0.352 0.2030.3240.1680.248–0.2510.746 SocSci210 NJ5DX0.192 0.366 0.2040.3310.1720.2470.3640.781– GSS–0.2040.3320.1490.2220.3350.6140.2130.574 OpinionQA W260.167 0.320–0.1510.2290.3200.6080.2050.533 OpinionQA W290.184 0.346 0.1950.315–0.3510.7600.2090.548 Table 1: Survey simulation results on the target survey datasets in the low data-access regime. The best-performing simulation approach for each target survey is indicated in bold, with the second-best underlined. As shown in Table 1, simulation performance varies across target survey datasets, partly reflect- ing differences in the average number of response options (SocSci210 subsets have an average of 7 response options and correspondingly result in highest TVD, whereas GSS has an average of 3.1 options and yields the lowest TVD). In addition, variation in survey topics may further contribute to performance differences. 9 Personas induced from diverse out-of-domain sources generally do not outperform simple base- lines that provide minimal context about target de- mographic groups. In particular, B3, which condi- tions each agent on a single attribute at inference time, remains consistently competitive. Notably, personas derived from sources such as shopping histories and movie reviews are also not particu- larly effective. One possible explanation is that per- sona sources often focus on relatively narrow topics (e.g., OpinionQA W29 centers on gender-related is- sues, while movie reviews reflect domain-specific preferences), leading to personas that are overly specific and misaligned with the target surveys. This suggests that, without strong alignment be- tween source behavioral data and the target popula- tion, additional persona detail may introduce noise rather than improve simulation alignment. Overall, the results indicate that group-level response distri- butions can often be approximated from minimal demographic context, and that naively incorporat- ing out-of-domain personas does not reliably add predictive signal. 9 TVD and WD generally produce similar method rankings; unless otherwise noted, we therefore base our analysis on TVD. These results should also be interpreted in light of what distribution alignment does not capture. Even when persona-conditioned simulation does not substantially improve simulation alignment, it can still offer practical advantages in interpretabil- ity and explainability. Because each simulated re- sponse is tied to a particular persona induced from behavioral data, the resulting predictions are eas- ier to inspect and explain than those from single- attribute baseline. Moreover, using a pool of per- sonas enables the simulator to capture variation within a demographic group, rather than represent- ing that group with a single-attribute demographic information. We attribute the limited gains from out-of- domain personas to two primary factors. First, behavioral data from a source dataset may con- tain signals that are not useful for simulating users’ opinions in target dataset. Second, our pipeline maps personas to target groups using demographic attributes inferred from behavioral data; inaccu- rate demographic attribute inference can therefore cause persona–group misassignment (e.g., assign- ing a persona derived from behavioral data of a female user to a male target demographic group because of incorrect demographic feature extrac- tion), degrading simulation alignment. 10 These observations motivate RQ2. 5.2 Mid-access Regime / Out-of-domain Setting with GT Demographics For datasets where individual-level GT demo- graphic attributes are available, we evaluate two 10 Details of demographic feature inference accuracy are provided in Tables 7 and 8 in Appendix F. Baseline/Persona Source GSSOpinionQA W26 OpinionQA W29 SocSci210 (MP3DR) SocSci210 (NJ5DX) TVDWDTVDWDTVDWDTVDWDTVDWD B30.151 0.280 0.1780.2870.1300.1940.3230.5830.2040.538 GSS–0.2040.3320.1490.2220.3350.6140.2130.574 +GT assignment–0.2000.3250.1410.2100.3190.5120.1940.474 +GT generation–0.2080.3360.1490.2360.317 0.4820.2000.495 OpinionQA W260.167 0.320–0.1510.2290.3200.6080.2050.533 +GT assignment0.1540.304–0.1410.2140.3050.5440.1920.472 +GT generation0.157 0.291 –0.1500.2340.3290.5710.1860.457 OpinionQA W290.184 0.346 0.1950.315–0.3510.7600.2090.548 +GT assignment0.159 0.304 0.172 0.274–0.3260.6480.1880.447 +GT generation0.159 0.293 0.1690.267–0.3480.6770.1870.450 Table 2: Survey simulation results on the target survey datasets in mid data-access regime. The best-performing ap- proach for each survey is indicated in bold, with the second-bestunderlined. GT assignment uses GT demographics only for persona–group assignment, while GT generation uses GT demographics during persona generation. strategies for leveraging GT to improve persona- conditioned survey simulation: (i) GT assignment, where GT demographic information is used only to improve the assignment of personas to target de- mographic groups, and (i) GT generation, where persona generation is conditioned on both behav- ioral data and the corresponding GT demographic information about the user. As shown in Table 2, access to GT demographics substantially improves persona-conditioned simu- lation. GT assignment consistently reduces TVD relative to using the same persona source without GT demographic information, indicating that demo- graphic attribute misassignment is a major contrib- utor to reduced cross-domain simulation alignment. In contrast, GT generation yields mixed effects: it can further reduce TVD in some settings, but it is not uniformly better than GT assignment, sug- gesting that conditioning persona generation on demographic information can introduce additional variance depending on the source behavioral data. Our qualitative analysis further suggests that con- ditioning persona generation on GT demograph- ics can make the resulting personas disproportion- ately focused on surface-level attributes, rather than higher-level traits such as interests, preferences, or behavioral tendencies (see Section H.1). Overall, these findings highlight accurate demographic as- signment as a key bottleneck: when personas are correctly assigned to target demographic groups, persona-conditioned simulation becomes substan- tially more aligned to human responses even when the source behavioral data differs from the target survey population. 20406080 % of questions used for persona generation 0.12 0.14 0.16 0.18 0.20 TVD In domain: GSS-based Personas In domain: GSS-based Personas + GT Assignment In domain: GSS-based Personas + GT Generation Baseline: B3 Figure 3: Survey simulation results in the in-domain setting, showing TVD for simulations using personas generated from varying fractions of GSS questions. 5.3 High-access Regime / In-domain Setting We investigate how well personas induced from target-survey data generalize within the same do- main (RQ3). We use the GSS dataset for this set- ting, as it covers diverse topics and includes enough questions to allow meaningful splits. Questions are divided into two subsets, with 20%, 40%, 60%, 80% used for persona generation and the remain- der for simulation evaluation. As shown in Figure 3, simulation performance improves as more question history is used, with the largest gains at 80%. One interpretation is that additional questions provide richer behavioral evidence, enabling the persona generator to infer more stable attributes, such as preferences or at- titudes toward topics related to the held-out ques- tions, which in turn improves response simulation. Moreover, when personas incorporate a larger por- tion of each respondent’s history, the simulator can better capture consistent response patterns across related questions. This leads to more faithful simu- lations of the respondent represented by each per- 0.10.20.30.40.50.6 Demographic mismatch (TVD) for target demographic attributes 0.10 0.15 0.20 0.25 0.30 0.35 TVD Target key AGE CREGION EDUCATION INCOME MARITAL POLIDEOLOGY POLPARTY RACE RELIG RELIGATTEND SEX Figure 4: For each target demographic group, we mea- sure the correlation between demographic mismatch score (as measured by TVD between selected personas and actual demographics) and survey TVD. Correlation coefficient: (Pearson=0.631, Spearman = 0.323) sona and, consequently, better alignment between human and agent response distributions within spe- cific demographic groups. In-domain personas with GT generation consis- tently outperforms the baseline once 40% or more of the question history is used. This result suggests that personas generalize effectively across different question sets within the same domain, and further indicates that access to a sufficient amount of be- havioral data from the target source is important for improving survey simulation. To verify that this finding is not an artifact of a single data split, we repeat the experiment across multiple splits. GT Generation outperforms B3 in all three independent runs at the 80% split (0.111 vs. 0.131; 0.139 vs. 0.147; 0.153 vs. 0.169). While marginal standard deviations overlap between con- ditions (see Figure 7 for full results including stan- dard deviations), this paired comparison confirms that the observed ranking is stable across splits. Due to computational cost, remaining conditions in Tables 1, 2, and 5 are run once; we treat this repeated-run result as verification that our pipeline produces reproducible relative orderings. 6 Discussion In this section, we provide deeper analyses of our persona-conditioned simulation approaches and baselines. Unless stated otherwise, all anal- yses adopt the GSS evaluation configuration in which personas are induced from 80% of the sur- vey questions and evaluated on the remaining 20% (as described in Section 5.3). We focus on rep- resentative methods from each investigated data- access regime spanning the single-attribute context Figure 5: TVD across different demographic-group sizes after binning groups into Small, Medium, and Large buckets by respondent count in GSS dataset. baseline (B3), model-knowledge personas (Model Knowledge), out-of-domain personas with GT as- signment sourced from OpinionQA Wave 26 (Opin- ionQA), and in-domain personas induced from 80% of the question history with GT demographic infor- mation (GSS 80%). 6.1 Effect of Persona Representativeness on Simulation Alignment We investigate whether simulation performance de- pends on how well the selected personas capture the ground-truth respondent population beyond the tar- get attribute. Under the GSS 80% setting, we com- pare, for each target subgroup (e.g.,Gender=male orAge=18–29), the 30 personas used for simula- tion with the corresponding respondent subgroup. For each non-target demographic attribute, we com- pute TVD between the attribute distribution across the selected personas and the distribution in the corresponding respondent group, considering all possible attribute values (e.g., education levels or income brackets). This yields a separate TVD score for each non-target attribute, which we then aver- age to obtain an overall demographic mismatch score for the demographic group. As shown in Figure 4, we observe a moder- ate positive association between demographic mis- match and simulation misalignment (TVD) (Pear- son= 0.631, Spearman= 0.323), suggesting that demographic groups for which selected personas more closely resemble the corresponding respon- dent population tend to exhibit lower simulation misalignment. The weaker Spearman correlation, however, indicates that the relationship is not uni- formly monotonic across demographic groups. AGE CREGION EDUCATION INCOME MARITAL POLIDEOLOGY POLPARTY RACE RELIG RELIGATTEND SEX Demographic Keys B3 Model Knowledge OpinionQA GSS 80% Methods 0.1160.1120.1210.1340.1450.1500.1230.1310.1510.1370.119 0.1270.1280.1240.1320.1490.1550.1540.1410.1440.1350.129 0.1400.1380.1370.1470.1560.1680.1490.1500.1570.1530.141 0.1080.1090.1050.1140.1220.1060.1150.1190.1140.1070.108 Performance Heatmap (TVD) by Demographic Keys 0.11 0.12 0.13 0.14 0.15 0.16 TVD Figure 6: Performance Heatmap (TVD) across demographic keys 6.2 Simulation Alignment Variations across Demographic Groups Group size. We first compare performance across demographic-group sizes to assess whether simulation remains effective for relatively under- represented groups across different data-access regimes. 11 As shown in Figure 5, the ranking of methods is stable across all group sizes: personas induced from 80% of GSS questions consistently achieve the best simulation alignment, followed by B3 and Model Knowledge, while OpinionQA exhibits the highest TVD. This suggests that the benefits of in-domain persona induction are not limited to well-populated groups but extend to both sparse and large groups. The advantage is especially pro- nounced for smaller groups, where underrepresen- tation presents a greater challenge. Overall, these results suggest that in-domain persona induction yields stronger and more stable alignment across groups of varying sizes. Demographic key. We next compare perfor- mance across demographic keys (e.g.,AGEand REGION) to examine which attributes are more con- sistently captured by each method. As shown in Figure 6, GSS 80% performs best across all demo- graphic attributes, with uniformly low TVD val- ues. This indicates that personas induced from in-domain evidence generalize reliably across at- tributes, rather than improving only a narrow subset of groups. Among the remaining methods, B3 is generally the strongest baseline, particularly on age, education, race, and sex, where differences are relatively small. By contrast, OpinionQA con- sistently underperforms, indicating limited cross- 11 The number of respondents in GSS does not fully reflect the true size of each demographic group, but it can serve as a proxy for population representation (e.g., smaller groups may be viewed as relatively underrepresented groups). domain simulation alignment when the source dis- tribution diverges from the target. Model Knowl- edge performs between B3 and OpinionQA, and overall shows notable degradation on POLPARTY (0.154), suggesting that political affiliation is espe- cially difficult to simulate from model knowledge alone when the target distribution differs from the model’s internal knowledge. 7 Conclusion We presented a systematic study of persona- conditioned survey simulation under different data- access regimes, evaluated at the demographic- group level. Our results show that out-of-domain personas rarely outperform a single-attribute con- text baseline, primarily due to domain mismatch and errors in demographic attribute assignment. Their effectiveness improves substantially when ground-truth demographics are available, partic- ularly for persona–group assignment. In-domain personas perform progressively better as more ques- tion history is incorporated, suggesting that richer behavioral evidence enables more stable trait in- ference and better generalization to held-out ques- tions. Overall, these findings highlight a practical insight for survey simulation: the critical factor is not merely the use of personas, but ensuring they are grounded in sufficient behavioral evidence and accurately matched to the target population. Limitations First, our evaluation has several scope limitations: it focuses on single-attribute demographic groups rather than intersectional groups, which may un- derstate the difficulty of simulating finer-grained subpopulations; it is limited to U.S. respondents be- cause our behavioral data sources are U.S.-centric; and it evaluates distribution-level alignment rather than individual-level, prioritizing privacy-aligned deployment. Extending the evaluation to intersec- tional groups and other regions, and connecting distributional accuracy to downstream decision- making and application-specific utility, remain im- portant directions for future work. Second, we did not investigate methods for im- proving demographic feature extraction accuracy. Table 8 shows that extraction errors are a major bottleneck in the low-access regime, as they lead to persona misassignment and degrade simulation alignment. Our mid-access results suggest that ac- curate demographic assignment is the key factor, providing a concrete target for future work on more robust demographics inference methods. Finally, due to computational cost, most experi- ments are run only once. Rather than focusing on stability through repeated runs, we prioritize gen- eralizability by evaluating across 5 target datasets, 8 persona sources, and 3 LLMs, where consistent trends are observed. Since this is the first work to characterize persona generation from diverse be- havioral data domains and systematically measure persona fidelity across domains, we prioritize gen- eralization over stability. For the in-domain setting, where uncertainty estimation is most relevant given the varying question history splits, we conduct mul- tiple runs and report standard deviation in Figure 7 and Table 6. Ethical Considerations In our experiments, we utilize publicly available datasets, including OpinionQA (Santurkar et al., 2023), GSS (Davern et al., 2024), SocSci210 (Kol- luri et al., 2025), movie review (Yoon et al., 2024), and shopping histories (Berke et al., 2024). These datasets are widely adopted and well-established within the research community, and their use raises no additional privacy or consent concerns beyond those addressed in the original dataset releases. All large language models used in this work were accessed through their official and publicly avail- able sources. Our use of these models complies with their stated terms of service and aligns with open science and reproducibility principles. During the preparation of this manuscript, we employed an AI-assisted writing tool at the sen- tence level to support drafting and linguistic refine- ment. Beyond the privacy safeguards already described, we note additional societal risks specific to persona- conditioned survey simulation. First, tools that simulate group-level opinion distributions could be misused for political microtargeting or opinion manipulation. Second, survey simulation risks be- ing used as a substitute for genuine stakeholder engagement, particularly for underrepresented or hard-to-reach populations. Our results suggest this is especially risky in the low-access regime, where simulated distributions can appear plausible while diverging substantially from real subgroup opinions due to demographic misassignment (Section 5.1). We emphasize that our findings argue against us- ing out-of-domain persona simulation as a drop-in replacement for genuine data collection, and rec- ommend that any deployment report demographic- assignment accuracy and group-level sample size alongside simulated distributions. References Gati V Aher, Rosa I Arriaga, and Adam Tauman Kalai. 2023. Using large language models to simulate mul- tiple humans and replicate human subject studies. In International conference on machine learning, pages 337–371. PMLR. Anthropic. 2025. Claude 3.7 sonnet and claude code. Lisa P Argyle, Ethan C Busby, Nancy Fulda, Joshua R Gubler, Christopher Rytting, and David Wingate. 2023. Out of one, many: Using language mod- els to simulate human samples. Political Analysis, 31(3):337–351. Christopher A Bail. 2024. Can generative ai improve so- cial science? Proceedings of the National Academy of Sciences, 121(21):e2314021121. Ziyad Benomar, Weronika Łajewska, Leonardo Perelli, and Saab Mansour. 2026.Data-driven persona- conditioned agents for A/B test simulation. arXiv preprint arXiv:2609.01038. Alex Berke, Dan Calacci, Robert Mahari, Takahiro Yabe, Kent Larson, and Sandy Pentland. 2024. Open e-commerce 1.0, five years of crowdsourced us ama- zon purchase histories with user demographics. Sci- entific Data, 11(1):491. James Bisbee, Joshua D Clinton, Cassy Dorff, Brenton Kenkel, and Jennifer M Larson. 2024. Synthetic replacements for human survey data? the perils of large language models. Political Analysis, 32(4):401– 416. Yong Cao, Haijiang Liu, Arnav Arora, Isabelle Augen- stein, Paul Röttger, and Daniel Hershcovich. 2025. Specializing large language models to simulate sur- vey response distributions for global populations. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Compu- tational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 3141–3154. Serina Chang, Alicja Chaszczewicz, Emma Wang, Maya Josifovska, Emma Pierson, and Jure Leskovec. 2025. Llms generate structurally realistic social net- works but overestimate political homophily. In Pro- ceedings of the International AAAI Conference on Web and Social Media, volume 19, pages 341–371. Jiangjie Chen, Xintao Wang, Rui Xu, Siyu Yuan, Yikai Zhang, Wei Shi, Jian Xie, Shuang Li, Ruihan Yang, Tinghui Zhu, and 1 others. 2024. From persona to personalization: A survey on role-playing language agents. arXiv preprint arXiv:2404.18231. Myra Cheng, Tiziano Piccardi, and Diyi Yang. 2023. Compost: Characterizing and evaluating caricature in llm simulations. In Proceedings of the 2023 Con- ference on Empirical Methods in Natural Language Processing, pages 10853–10875. Michael Davern, Rene Bautista, Jeremy Freese, Pamela Herd, and Stephen L. Morgan. 2024. General social survey 2024. [Machine-readable data file]. Principal Investigator: Michael Davern; Co-Principal Investi- gators: Rene Bautista, Jeremy Freese, Pamela Herd, and Stephen L. Morgan. Sponsored by National Sci- ence Foundation. Data accessed from the GSS Data Explorer. Ricardo Dominguez-Olmedo, Moritz Hardt, and Celes- tine Mendler-Dünner. 2024. Questioning the survey responses of large language models. Advances in Neural Information Processing Systems, 37:45850– 45878. Chen Gao, Xiaochong Lan, Nian Li, Yuan Yuan, Jingtao Ding, Zhilun Zhou, Fengli Xu, and Yong Li. 2024. Large language models empowered agent-based mod- eling and simulation: A survey and perspectives. Humanities and Social Sciences Communications, 11(1):1259. Tao Ge, Xin Chan, Xiaoyang Wang, Dian Yu, Haitao Mi, and Dong Yu. 2024. Scaling synthetic data cre- ation with 1,000,000,000 personas. arXiv preprint arXiv:2406.20094. Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al- Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, and 1 others. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783. Luke Hewitt, Ashwini Ashokkumar, Isaias Ghezae, and Robb Willer. 2024. Predicting results of social science experiments using large language models. Preprint. John J Horton. 2023. Large language models as sim- ulated economic agents: What can we learn from homo silicus? Technical report, National Bureau of Economic Research. Tiancheng Hu, Joachim Baumann, Lorenzo Lupo, Nigel Collier, Dirk Hovy, and Paul Röttger. 2026. Sim- bench: Benchmarking the ability of large language models to simulate human behaviors. In Interna- tional Conference on Learning Representations, vol- ume 2026, pages 28290–28341. EunJeong Hwang, Bodhisattwa Majumder, and Niket Tandon. 2023. Aligning language models to user opinions. In Findings of the Association for Com- putational Linguistics: EMNLP 2023, pages 5906– 5919. Shivani Kapania, William Agnew, Motahhare Eslami, Hoda Heidari, and Sarah E Fox. 2025. Simulacrum of stories: Examining large language models as qual- itative research participants. In Proceedings of the 2025 CHI Conference on Human Factors in Comput- ing Systems, pages 1–17. Jaehyung Kim and Yiming Yang. 2025. Few-shot per- sonalization of llms with mis-aligned responses. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Compu- tational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 11943–11974. Junsol Kim and Byungkyu Lee. 2023. Ai-augmented surveys: Leveraging large language models and surveys for opinion prediction.arXiv preprint arXiv:2305.09620. Akaash Kolluri, Shengguang Wu, Joon Sung Park, and Michael S Bernstein. 2025. Finetuning llms for hu- man behavior prediction in social science experi- ments. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 30084–30099. Louis Kwok, Michal Bravansky, and Lewis D Griffin. 2024. Evaluating cultural adaptability of a large lan- guage model via simulation of synthetic personas. arXiv preprint arXiv:2408.06929. Weronika Łajewska, Paul Missault, George Davidson, and Saab Mansour. 2026. RADIUS: Ranking, dis- tribution, and significance-a comprehensive align- ment suite for survey simulation. arXiv preprint arXiv:2603.19002. Benjamin S Manning, Kehang Zhu, and John J Horton. 2024. Automated social science: Language models as scientist and subjects. Technical report, National Bureau of Economic Research. Saab Mansour, Leonardo Perelli, Lorenzo Mainetti, George Davidson, and Stefano D’Amato. 2025. Paars: Persona aligned agentic retail shoppers. In Proceedings of the 1st Workshop for Research on Agent Language Models (REALM 2025), pages 143– 159. Nicole Meister, Carlos Guestrin, and Tatsunori B Hashimoto. 2025. Benchmarking distributional align- ment of large language models. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Lin- guistics: Human Language Technologies (Volume 1: Long Papers), pages 24–49. Suhong Moon, Marwa Abdulhai, Minwoo Kang, Joseph Suh, Widyadewi Soedarmadji, Eran Kohen Behar, and David M Chan. 2024. Virtual personas for lan- guage models via an anthology of backstories. In Pro- ceedings of the 2024 conference on empirical meth- ods in natural language processing, pages 19864– 19897. Joon Sung Park, Carolyn Q Zou, Aaron Shaw, Ben- jamin Mako Hill, Carrie Cai, Meredith Ringel Morris, Robb Willer, Percy Liang, and Michael S Bernstein. 2024a. Generative agent simulations of 1,000 people. arXiv preprint arXiv:2411.10109. Peter S Park, Philipp Schoenegger, and Chongyang Zhu. 2024b. Diminished diversity-of-thought in a standard large language model. Behavior Research Methods, 56(6):5754–5770. Shibani Santurkar, Esin Durmus, Faisal Ladhak, Cinoo Lee, Percy Liang, and Tatsunori Hashimoto. 2023. Whose opinions do language models reflect? In In- ternational Conference on Machine Learning, pages 29971–30004. PMLR. Gabriel Simmons. 2023. Moral mimicry: Large lan- guage models produce moral rationalizations tailored to political identity. In Proceedings of the 61st An- nual Meeting of the Association for Computational Linguistics (Volume 4: Student Research Workshop), pages 282–297. Joseph Suh, Erfan Jahanparast, Suhong Moon, Minwoo Kang, and Serina Chang. 2025. Language model fine-tuning on scaled survey data for predicting dis- tributions of public opinions. In Proceedings of the 63rd Annual Meeting of the Association for Compu- tational Linguistics (Volume 1: Long Papers), pages 21147–21170. Chenkai Sun, Ke Yang, Revanth Gangi Reddy, Yi R Fung, Hou Pong Chan, Kevin Small, Cheng Xiang Zhai, and Heng Ji. 2025. Persona-db: Efficient large language model personalization for response predic- tion with collaborative data refinement. In 31st Inter- national Conference on Computational Linguistics, COLING 2025, pages 281–296. Association for Com- putational Linguistics (ACL). Seungjong Sun, Eungu Lee, Dongyan Nan, Xiangy- ing Zhao, Wonbyung Lee, Bernard J Jansen, and Jang Hyun Kim. 2024.Random silicon sam- pling: Simulating human sub-population opinion using a large language model based on group- level demographic information.arXiv preprint arXiv:2402.18144. Yu-Min Tseng, Yu-Chao Huang, Teng-Yun Hsiao, Wei- Lin Chen, Chao-Wei Huang, Yu Meng, and Yun- Nung Chen. 2024. Two tales of persona in llms: A survey of role-playing and personalization. In Find- ings of the Association for Computational Linguistics: EMNLP 2024, pages 16612–16631. Lei Wang, Jingsen Zhang, Hao Yang, Zhi-Yuan Chen, Jiakai Tang, Zeyu Zhang, Xu Chen, Yankai Lin, Hao Sun, Ruihua Song, and 1 others. 2025. User behavior simulation with large language model-based agents. ACM Transactions on Information Systems, 43(2):1– 37. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, and 1 others. 2025.Qwen3 technical report.arXiv preprint arXiv:2505.09388. Se-eun Yoon, Zhankui He, Jessica Echterhoff, and Ju- lian McAuley. 2024. Evaluating large language mod- els as generative user simulators for conversational recommendation. In Proceedings of the 2024 Con- ference of the North American Chapter of the Asso- ciation for Computational Linguistics: Human Lan- guage Technologies (Volume 1: Long Papers), pages 1490–1504. Persona Source DatasetDomain#Q Avg. Resp. #Resp. GSSGeneral Survey6036.13309 OpinionQA W26Survey: Guns7846.24168 OpinionQA W29Survey: Gender7753.84867 SocSci210 MP3DR Survey: Gender396.94161 SocSci210 NJ5DXSurvey: Inequality4818.21814 IMDBMovie review–21.21079 E-commerceShopping history–368.25027 Test Dataset Dataset#Demog. feats.#Q#Opt.Total GSS40603.12280 OpinionQA W2650784.53474 OpinionQA W2950774.13622 SocSci210 MP3DR 18397.0555 SocSci210 NJ5DX25447.0934 Table 3: Summary statistics of persona sources and test datasets. Figure 7: Survey simulation results in the in-domain setting, showing TVD and standard deviation (Std.) for simulations using personas generated from varying frac- tions of GSS questions. Results for each iteration and Std. are available in Table 6 A Pilot Study In our pilot study, we explore several design choices for using synthetic personas in survey- response simulation. A.1 Verbalized Distribution vs. Option Selection Section 3.1 elicits a verbalized probability distri- bution from each persona-conditioned agent (e.g., A: 30%, B: 70%) and then averages these dis- tributions. As an alternative, we can estimate the group-level distribution by prompting each agent to select a single answer option from the available set and then aggregating these selections across agents. Concretely, for each agentj ∈1, . . . , N A con- ditioned on personaP j , we prompt the LLM to output exactly one optionˆx (j) q ∈X q for questionq. We then form an empirical distribution by counting Figure 8: Performance of Claude-3.7-Sonnet and Llama- 3.1-70B model on OpinionQA across different number of personas and different distribution aggregation meth- ods how often each option is selected: S q,g (x) = 1 N A N A X j=1 1 n ˆx (j) q = x o , x∈X q , (1) where 1· denotes the indicator function. A.2 Number of Personas and Aggregation Scheme: Verbalized Distributions vs. Option Selection We compare two aggregation schemes: (i) verbal- ized distribution outputs, in which the simulator directly produces a probability distribution over answer options, and (i) option selection + aggre- gation, in which each persona selects a single op- tion and responses are aggregated by voting. We also vary the number of personas used for each demographic group. We conduct experiments on SimBench- OpinionQA (Hu et al., 2026) using Claude 3.7 Sonnet and Llama-3.1-70B as simulators. For each target demographic value, we evaluate simulations using 30, 50, and 100 personas, and compare the two aggregation schemes. As shown in Figure 8, verbalized distributions consistently outperform option selection + aggrega- tion. This finding aligns with prior work suggesting that direct distribution elicitation better preserves minority probability mass than hard voting (Meis- ter et al., 2025). We also find that performance largely saturates beyond 30 personas: once the number of personas exceeds approximately 30, the improvement curve plateaus, suggesting diminish- ing returns from adding more agents under this setup. A.3 Why Does Option Selection Underperform? We next investigate why the option selection scheme performs worse than verbalized distri- GT GenerationFree Youarea33-year-oldBlackmanlivingintheNortheast whoworksineducationadministrationatanelementaryor secondaryschool.You’reacollegegraduatewitha bachelor’sdegreewhohasnevermarriedbuthastwo children.YouidentifyasProtestantandattendreligious servicesweekly,indicatingthatfaithisanimportantpartof yourlife. Politically, you lean Democratic but consider yourself moderate rather than strongly liberal. You consistently vote in presidential elections, supporting Democratic candidates like Clinton and Biden. You have only some confidence in the executive branch of government, suggesting a degree of political skepticism despite your party affiliation... You are a politically moderate Democrat who consistently votes in elections and supports Democratic presidential candidates. While you identify as middle-of-the-road politically, your policy positions tend to lean progressive on domestic issues - you support increased spending on education, healthcare, environmental protection, addressing urban problems, and improving conditions for Black Americans. You oppose the death penalty and believe premarital sex is not morally wrong. You’re fiscally more conservative on international matters, believing we spend too much on military and foreign aid. You’re civically engaged and vote regularly in presidential elections... Table 4: GT feature vs. Free description Figure 9: Average Shannon entropy of simulated re- sponse distributions under the optionselection aggre- gation scheme. Lower entropy indicates more peaked (less diverse)distributions, showing that option selection tends to collapse onto a single dominant option. butions. Our working hypothesis is that major- ity voting over single-choice responses produces overly peaked distributions that over-represent the dominant preference while systematically under- representing minority opinions. To examine this hypothesis, we compute two diagnostics: (1) the average Shannon entropy of the simulated response distribution, where lower entropy indicates a more concentrated distribution, and (2) the number of non-chosen options across questions, where for each of the 964 survey ques- tions we count how many answer options receive zero selections under option selection + aggrega- tion. Across questions, we find that option selection frequently collapses to a single dominant option, with most of the probability mass concentrated on one answer and many alternatives never selected even once. As a result, the simulated distribution primarily reflects the majority view and fails to capture the within-group heterogeneity observed in human data. Figure 10: Number of answer options with zero counts under the option selection aggregation scheme, aggre- gated across survey questions. A high number of zero- count options indicates that minority opinions are rarely represented in the simulated distributions. This effect is further amplified by the persona construction used in this pilot setting: personas are generated from only a single demographic attribute (e.g., Gender: female). With such limited informa- tion to differentiate personas, agents tend to con- verge on the same “typical” response, and voting- based aggregation then amplifies this consensus while suppressing minority mass. This mechanism plausibly explains why option-selection-based dis- tributions align less well with the true subgroup- level response distributions. A.4 Persona Induction Strategies To study how prompting affects persona quality, we compare four induction strategies: 1.Free. Generate a free-form bio from the indi- vidual’s behavioral data only. 2.Free+GT. Generate a free-form bio from the individual’s behavioral data + Ground- truth demographic feature as described in Sec- tion 5.2. 3.Target-aware. Provide brief context about the target survey datasets to be simulated (task description, domains, and example questions), then generate a bio. 4.Rubric-based. Ask the generator to answer survey-style rubric questions (aligned with tar- get survey domains) based on the individual’s data, then synthesize a bio. Each persona includes (i) a short bio and (i) extracted demographic features with confidence scores. Personas are then grouped by demographic features (e.g., Gender: male) and selected accord- ingly for subgroup simulation. In pilot experiments, we found that using the Free prompt template for persona generation and selecting personas with high confidence on the target demographic attribute consistently yielded the best performance. Detailed prompt templates used in the experiments (Free, Free+GT) are described in Figure 11 and 12, re- spectively. B Persona Detail Each key has a closed set of possible values: •Region: closed options (South | Northeast | West | Midwest) •Age: closed options (18–29 | 30–49 | 50–64 | 65+) • Sex: closed options (male | female) •Education: closed options (high school grad- uate | college graduate/some postgrad | some college, no degree | associate’s degree | post- graduate | less than high school) • Marital: closed options (married | divorced | widowed | living with a partner | never been married) •Religion: closed options (protestant | roman catholic | atheist | agnostic | nothing in partic- ular) •Religious Attendance: closed options (more than once a week | seldom | never | a few times a year | once or twice a month | once a week) • Political Party: closed options (Democrat | Independent | Other | Republican) •Income: closed options (Less than $30,000 | $30,000–$50,000 | $50,000–$75,000 | $75,000–$100,000 | $100,000 or more) •Political Ideology: closed options (liberal | moderate | conservative | very conservative | very liberal) •Race: closed options (black | hispanic | asian | white) C Metric Details TVD measures theℓ 1 discrepancy between two discrete distributions: TVD(U i,g , S i,g ) = 1 2 N option,i X k=1 |U i,g (x k )− S i,g (x k )| . (2) WD captures the minimal “mass transport” needed to transform one distribution into another while respecting the ordering of options. 12 W 1 (U i,g , S i,g ) = N option,i X k=1 k X r=1 U i,g (x r )− k X r=1 S i,g (x r ) . (3) D Dataset Details D.1 Statistics of Target Datasets and Persona-source Datasets As discussed in Section 4, we utilize 5 different survey datasets for test and use various sources of persona induction. Table 3 summarize the statistics for datasets used in our experiments. D.2 Group-level Test Instances Example Figure 13 shows example test instances from each target survey datasets. E Inference Details For persona generation we used temperature = 0.7 and for simulation inference (both baseline and per- sona conditioned inference) we used temperature = 0.0 Prompt used for synthetic persona-conditioned simulation is described in Figure 14. 12 This metric complements TVD by accounting for the ordinal structure of response options. All target datasets used in our experiments consist of survey questions with ordinal options, which are suited for WD. F Accuracy of Demographic Feature Extraction in Persona Generation In this section, we evaluate whether demographic features extracted from behavior data match ground-truth demographics. Ground-truth demo- graphics are available for OpinionQA W26/W29 and GSS. F.1 Extraction Accuracy by Confidence Level We first analyze how extraction accuracy varies with the model-assigned confidence level. As shown in Table 7, accuracy usually increases with confidence across datasets, consistent with Ap- pendix A’s finding that high-confidence persona selection improves simulation performance. We then test whether the personas used for sim- ulation (N A = 30per demographic feature) are actually matched to the intended target group. The results shown in Table 8 indicate substantial mis- match, which helps explain the weaker perfor- mance of personas constructed without ground- truth demographic conditioning: when persona demographics are misaligned with the target sub- group, persona conditioning can become noisy or misleading. G Additional Results To test robustness of survey simulation across model families, we report results for two additional models (Llama-3.3-70B (Grattafiori et al., 2024) and Qwen-3-32B (Yang et al., 2025)) using per- sonas induced from diverse sources in Table 5. The findings in Section 5.1 regarding persona utility hold consistently across models. H Additional Discussions H.1 Personas Induced from Behavioral Data Alone vs. with GT Demographics Providing GT demographics during persona genera- tion sometimes reduces TVD in certain settings, but it does not uniformly outperform GT assignment, as discussed in Section 5.2. Our manual inspection of these personas suggests a systematic difference in how they are constructed. Personas generated with GT demographics typically begin with an ex- plicit listing of the provided demographic attributes, followed by a behavioral summary grounded in the respondent’s observed answers. In contrast, per- sonas generated without GT demographic informa- tion often express demographic characteristics only implicitly, or leave them underspecified, while re- lying more heavily on generic behavioral patterns (see Appendix Table 4 for an example). These observations suggest that providing GT demographic information may encourage the gen- eration of shallower personas that focus less on actual behavioral patterns. This may help explain why GT-informed persona generation does not con- sistently lead to improved simulation performance. H.2 Broader Implications for Survey Simulation Systems Taken together, our results suggest that the main value of synthetic personas for survey simulation is not simply that they provide richer prompt text, but that they can encode behaviorally grounded evidence when they are representative of the pop- ulation being simulated. In practice, this means future survey simulation systems should treat per- sona use as an alignment problem rather than a prompting problem: they should validate how well personas represent source-population and explicitly audit persona–group matching. Our findings also suggest that when GT demographics are available, they are especially useful for matching, whereas injecting them directly into persona generation may encourage shallower, surface-level profiles. Finally, when partial target-survey history is available, sys- tems should prioritize it, since richer in-domain behavioral evidence yields more stable traits, better transfer to unseen questions, and more consistent gains across demographic groups. The practical impact of this work is therefore to provide design guidance for building more reliable and privacy- aligned survey simulators using synthetic personas derived from diverse behavioral data. Simulator: Claude 3.7 Target DataGSSOpinionQA W26 OpinionQA W29 SocSci210 (MP3DR) SocSci210 (NJ5DX) TVDWDTVDWDTVDWDTVDWDTVDWD B10.169 0.3220.1890.2990.1710.2620.3110.4740.2410.678 B2 0.1600.3010.1730.2740.1390.2040.3140.5520.2040.560 B30.151 0.2800.1780.2870.1300.1940.3230.5830.2040.538 Persona Source Model Knowledge0.1600.3030.1900.3110.1570.2370.3300.5790.1910.443 Movie0.169 0.3260.1970.3180.1760.2630.3150.5920.2030.506 Shopping His0.163 0.3120.1890.3010.1600.2400.3090.5320.2080.530 Survey MP3DR 0.186 0.3520.2030.3240.1680.248–0.2510.746 Survey NJ5DX0.192 0.3660.2040.3310.1720.2470.3640.781– OpinionQA W26 0.167 0.320–0.1510.2290.3200.6080.2050.533 OpinionQA W290.184 0.3460.1950.315–0.3510.7600.2090.548 GSS –0.2040.3320.1490.2220.3350.6140.2130.574 Simulator: Llama-3.3-70B Target DataGSSOpinionQA W26 OpinionQA W29 SocSci210 (MP3DR) SocSci210 (NJ5DX) TVDWDTVDWDTVDWDTVDWDTVDWD B10.200 0.3570.2000.3180.2000.2950.4750.8630.3050.879 B20.168 0.3200.1830.2870.1880.2790.4560.8610.3251.030 B30.171 0.3110.1830.2910.1810.2640.4310.8080.3110.947 Persona Source Model Knowledge0.1690.3200.1850.2970.1950.2940.4270.6870.2750.803 Movie0.191 0.3620.2100.3410.2390.3660.4720.8680.3150.983 Shopping His0.180 0.3420.2120.3450.2080.3170.4410.8200.2770.864 Survey MP3DR0.209 0.3920.1990.3220.2350.348–0.3901.290 Survey NJ5DX0.209 0.3980.2000.3230.2250.3340.3980.787– OpinionQA W260.192 0.355–0.2080.3090.4530.8320.2830.860 OpinionQA W290.194 0.3660.2040.327–0.3840.6600.2750.835 GSS–0.2090.3360.1980.3060.4310.7820.3030.943 Simulator: Qwen3-32B Target DataGSSOpinionQA W26 OpinionQA W29 SocSci210 (MP3DR) SocSci210 (NJ5DX) TVDWDTVDWDTVDWDTVDWDTVDWD B10.186 0.3370.2280.3810.2270.3500.5951.6740.2470.710 B20.180 0.3280.2350.3900.1890.2770.5561.4730.2600.721 B30.178 0.3190.2210.3690.1940.2820.5471.4860.2540.716 Persona Source Model Knowledge0.173 0.3110.2070.3460.1920.2840.5591.4510.2190.561 Movie0.183 0.3240.2290.3800.2140.3210.5421.2450.2530.699 Shopping His0.1740.3140.2180.3580.1930.2820.5231.3060.2260.626 Survey MP3DR0.199 0.3660.2280.3760.2120.310–0.2840.846 Survey NJ5DX0.203 0.3720.2220.3730.2170.3150.5921.594– OpinionQA W260.176 0.323–0.1920.2830.4921.2000.2290.621 OpinionQA W290.184 0.3410.2170.363–0.6091.7380.2180.586 GSS–0.2150.3610.1980.2910.5351.2540.2290.643 Table 5: Survey simulation results using three different LLMs on the target survey datasets in the low data-access regime. The best-performing simulation approach for each target survey is indicated in bold, with the second-best underlined. IterationQuestion SplitGSS Persona+GT GenerationB3 1 20%0.1870.1630.157 40%0.2020.1660.167 60%0.1780.1580.156 80%0.1410.1110.131 2 20%0.1780.1540.149 40%0.1730.1510.150 60%0.1830.1580.159 80%0.1690.1390.147 3 20%0.1860.1460.152 40%0.1810.1540.160 60%0.1890.1630.171 80%0.1780.1530.169 Mean± STD 20%0.184± 0.005 0.154± 0.009 0.153± 0.004 40%0.185± 0.015 0.157± 0.008 0.159± 0.009 60%0.183± 0.006 0.160± 0.003 0.162± 0.008 80%0.163± 0.019 0.134± 0.021 0.149± 0.019 Table 6: Survey simulation results in the in-domain setting across iterations, showing TVD for simulations using personas generated from varying fractions of GSS questions. Prompt for Persona Generation (Free) Role Definition (System Prompt): You are an AI assistant specialized in creating simulation-ready personas from survey responses. These personas will be used to simulate how the same individual would answer NEW survey questions. PRIMARY OBJECTIVE: Create a persona that maximizes predictive fidelity for future survey responses across politics, values, morality, institutions, lifestyle choices, and social issues. KEY PRINCIPLES: - Think holistically: infer stable traits that drive survey answers (ideology, religiosity, trust, moral reasoning style, risk tolerance, etc.). - Look for patterns: consistency across items implies stable stances; mixed answers imply conditional rules. - Build coherence: your persona must form a realistic, internally consistent person. - Distinguish evidence levels: separate direct evidence from educated inference. - Avoid fabrication: do not invent vivid biographical specifics unless directly supported. CRITICAL REQUIREMENTS: 1) Output MUST be valid JSON and MUST contain EXACTLY four top-level keys: "step1_analysis", "short_bio", "step2_analysis", "user_profile" Do NOT output any other keys. 2) SECOND-PERSON ONLY + no name: - The "short_bio" MUST be written in SECOND-PERSON perspective and refer to the person as "you". - Do NOT use any specific name. Address the person directly as "you". 3) "short_bio" is NOT a short biography. It is FREE-FORM persona_text for simulation: - No length constraint. Include all details you judge useful for downstream survey simulation. - It does NOT have to mention lifestyle/life stage/interests. - Focus on predictive traits, stances, conditional rules, attitude strength, and likely answer tendencies. 4) You MUST extract ALL 11 REQUIRED demographic fields: - CREGION, AGE, SEX, EDUCATION, MARITAL, RELIG, RELIGATTEND, POLPARTY, POLIDEOLOGY, INCOME, RACE - For each field, "value" MUST be selected ONLY from the provided options. - Provide "reasoning" BEFORE the value decision and include a "confidence" (high/medium/low). 5) No fabricated specifics: - Do NOT invent a job title, city, named organizations, number of children, or life events unless explicitly supported by input. - If you include an inference, keep it general and lower confidence accordingly. 6) You may include additional characteristics in "user_profile" beyond the 11 required fields, if they improve simulation fidelity. - For any non-required extra fields, "value" may be free-form, but must remain grounded in evidence and avoid fabricated specifics. IMPORTANT ABOUT ANALYSIS TEXT: - "step1_analysis" and "step2_analysis" should be evidence-based and helpful. - Cite specific survey questions (e.g., Q31) when possible. """ User Prompt: The user’s data is provided within <user_data></user_data> XML tags. <user_data> <behavior_data> behavior_data </behavior_data> </user_data> Please create a simulation-ready persona profile based on the survey responses provided. This persona will be used to predict how you (the individual) would respond to additional social surveys. Your task will be completed in TWO STEPS: === STEP 1: Induce Persona Text (Reasoning + Persona) === A) In "step1_analysis", analyze the survey responses holistically to understand you as a person whose answers can be simulated. Focus on: - Stable value system and worldview - Political/social attitudes and conditional rules (when answers vary by scenario) - Moral reasoning style (absolutist vs situational), risk tolerance, deference vs skepticism - Trust across institutions - Any life-stage/lifestyle indicators ONLY if supported by evidence (avoid fabrication) - What you would likely do when uncertain (default patterns) B) In "short_bio", write the persona_text itself (FREE-FORM, no length constraint): - SECOND-PERSON perspective ONLY (refer to the person as "you") - Do NOT use any specific name; address the person directly as "you" - NOT required to be a biography; can be bullets, paragraphs, or mixed format - Include all information you judge useful for accurate downstream survey simulation - Emphasize predictive traits, stance strength, and decision rules - Avoid invented specifics (no job/city/children counts unless explicitly supported) === STEP 2: Extract Required Fields (Reasoning + Structured Profile) === In "step2_analysis", explain how the survey responses AND the persona_text ("short_bio") support each inferred characteristic. In "user_profile", you MUST determine the following 11 REQUIRED fields, selecting values ONLY from the options: <profile_fields> - CREGION: User’s region in the US [Midwest, Northeast, South, West] - AGE: User’s age group [18-29, 30-49, 50-64, 65+] - SEX: User’s gender [Female, Male, In some other way] - EDUCATION: Education level [Less than high school, High school graduate, Some college, no degree, Associate’s degree, College graduate/some postgrad, Postgraduate] - MARITAL: Marital status [Married, Living with a partner, Divorced, Separated, Widowed, Never been married] - RELIG: Religious affiliation [Protestant, Roman Catholic, Mormon, Orthodox, Jewish, Muslim, Buddhist, Hindu, Atheist, Agnostic, Nothing in particular, Christian, Unitarian, Other] - RELIGATTEND: Religious service attendance [More than once a week, Once a week, Once or twice a month, A few times a year, Seldom, Never] - POLPARTY: Political party affiliation [Democrat, Republican, Independent, Other] - POLIDEOLOGY: Political ideology [Very liberal, Liberal, Moderate, Conservative, Very conservative] - INCOME: Annual household income [Less than 30, 000,30,000-50, 000,50,000-75, 000,75,000-100, 000,100,000 or more] - RACE: Racial/ethnic background [White, Black, Asian, Hispanic, Mixed Race, Other] </profile_fields> For each REQUIRED field in "user_profile", you MUST include: - "reasoning": evidence-based reasoning FIRST (cite question IDs where possible) - "value": pick ONLY from the provided options - "confidence": high/medium/low You MAY include additional characteristics beyond these 11 required fields (optional), but keep them grounded and avoid invented specifics. For extra fields, "value" may be free-form. === OUTPUT FORMAT (STRICT) === Output MUST be valid JSON with EXACTLY these four keys: "step1_analysis": "...", "short_bio": "...", "step2_analysis": "...", "user_profile":"CHARACTERISTIC_NAME": "reasoning": "...", "value": "...", "confidence": "high/medium/low" IMPORTANT: - Output JSON only. No extra text. - "short_bio" MUST be second-person ("you") and MUST NOT use any specific name. Figure 11: Prompt template used for Persona Generation (Free) as described in Appendix A.4. Prompt for Persona Generation (Free+GT) Role Definition (System Prompt): You are an AI assistant specialized in creating simulation-ready personas from survey responses. These personas will be used to simulate how the same individual would answer NEW survey questions. PRIMARY OBJECTIVE: Create a persona that maximizes predictive fidelity for future survey responses across politics, values, morality, institutions, lifestyle choices, and social issues. KEY PRINCIPLES: - Think holistically: infer stable traits that drive survey answers (ideology, religiosity, trust, moral reasoning style, risk tolerance, etc.). - Look for patterns: consistency across items implies stable stances; mixed answers imply conditional rules. - Build coherence: your persona must form a realistic, internally consistent person. - Distinguish evidence levels: separate direct evidence from educated inference. - Avoid fabrication: do not invent vivid biographical specifics unless directly supported. CRITICAL REQUIREMENTS: 1) Output MUST be valid JSON and MUST contain EXACTLY four top-level keys: "step1_analysis", "short_bio", "step2_analysis", "user_profile" Do NOT output any other keys. 2) SECOND-PERSON ONLY + no name: - The "short_bio" MUST be written in SECOND-PERSON perspective and refer to the person as "you". - Do NOT use any specific name. Address the person directly as "you". 3) "short_bio" is NOT a short biography. It is FREE-FORM persona_text for simulation: - No length constraint. Include all details you judge useful for downstream survey simulation. - It does NOT have to mention lifestyle/life stage/interests. - Focus on predictive traits, stances, conditional rules, attitude strength, and likely answer tendencies. 4) You MUST extract ALL 11 REQUIRED demographic fields: - CREGION, AGE, SEX, EDUCATION, MARITAL, RELIG, RELIGATTEND, POLPARTY, POLIDEOLOGY, INCOME, RACE - For each field, "value" MUST be selected ONLY from the provided options. - Provide "reasoning" BEFORE the value decision and include a "confidence" (high/medium/low). 5) No fabricated specifics: - Do NOT invent a job title, city, named organizations, number of children, or life events unless explicitly supported by input. - If you include an inference, keep it general and lower confidence accordingly. 6) You may include additional characteristics in "user_profile" beyond the 11 required fields, if they improve simulation fidelity. - For any non-required extra fields, "value" may be free-form, but must remain grounded in evidence and avoid fabricated specifics. IMPORTANT ABOUT ANALYSIS TEXT: - "step1_analysis" and "step2_analysis" should be evidence-based and helpful. - Cite specific survey questions (e.g., Q31) when possible. """ User Prompt: The user’s data is provided within <user_data></user_data> XML tags. <user_data> <features> features </features> <behavior_data> behavior_data </behavior_data> </user_data> Please create a simulation-ready persona profile based on the survey responses provided. This persona will be used to predict how you (the individual) would respond to additional social surveys. Your task will be completed in TWO STEPS: === STEP 1: Induce Persona Text (Reasoning + Persona) === A) In "step1_analysis", analyze the survey responses holistically to understand you as a person whose answers can be simulated. Focus on: - Stable value system and worldview - Political/social attitudes and conditional rules (when answers vary by scenario) - Moral reasoning style (absolutist vs situational), risk tolerance, deference vs skepticism - Trust across institutions - Any life-stage/lifestyle indicators ONLY if supported by evidence (avoid fabrication) - What you would likely do when uncertain (default patterns) B) In "short_bio", write the persona_text itself (FREE-FORM, no length constraint): - SECOND-PERSON perspective ONLY (refer to the person as "you") - Do NOT use any specific name; address the person directly as "you" - NOT required to be a biography; can be bullets, paragraphs, or mixed format - Include all information you judge useful for accurate downstream survey simulation - Emphasize predictive traits, stance strength, and decision rules - Avoid invented specifics (no job/city/children counts unless explicitly supported) === STEP 2: Extract Required Fields (Reasoning + Structured Profile) === In "step2_analysis", explain how the survey responses AND the persona_text ("short_bio") support each inferred characteristic. In "user_profile", you MUST determine the following 11 REQUIRED fields, selecting values ONLY from the options: <profile_fields> - CREGION: User’s region in the US [Midwest, Northeast, South, West] - AGE: User’s age group [18-29, 30-49, 50-64, 65+] - SEX: User’s gender [Female, Male, In some other way] - EDUCATION: Education level [Less than high school, High school graduate, Some college, no degree, Associate’s degree, College graduate/some postgrad, Postgraduate] - MARITAL: Marital status [Married, Living with a partner, Divorced, Separated, Widowed, Never been married] - RELIG: Religious affiliation [Protestant, Roman Catholic, Mormon, Orthodox, Jewish, Muslim, Buddhist, Hindu, Atheist, Agnostic, Nothing in particular, Christian, Unitarian, Other] - RELIGATTEND: Religious service attendance [More than once a week, Once a week, Once or twice a month, A few times a year, Seldom, Never] - POLPARTY: Political party affiliation [Democrat, Republican, Independent, Other] - POLIDEOLOGY: Political ideology [Very liberal, Liberal, Moderate, Conservative, Very conservative] - INCOME: Annual household income [Less than 30, 000,30,000-50, 000,50,000-75, 000,75,000-100, 000,100,000 or more] - RACE: Racial/ethnic background [White, Black, Asian, Hispanic, Mixed Race, Other] </profile_fields> For each REQUIRED field in "user_profile", you MUST include: - "reasoning": evidence-based reasoning FIRST (cite question IDs where possible) - "value": pick ONLY from the provided options - "confidence": high/medium/low You MAY include additional characteristics beyond these 11 required fields (optional), but keep them grounded and avoid invented specifics. For extra fields, "value" may be free-form. === OUTPUT FORMAT (STRICT) === Output MUST be valid JSON with EXACTLY these four keys: "step1_analysis": "...", "short_bio": "...", "step2_analysis": "...", "user_profile":"CHARACTERISTIC_NAME": "reasoning": "...", "value": "...", "confidence": "high/medium/low" IMPORTANT: - Output JSON only. No extra text. - "short_bio" MUST be second-person ("you") and MUST NOT use any specific name.""" Figure 12: Prompt template used for Persona Generation (Free+GT) as described in Appendix A.4. Target Test Instance Example GSS Question (q i ): Do you favor or oppose the death penalty for persons convicted of murder? Demographic group (g): ’REGION’: ’Midwest’ Answer Options (x k ): ’A’: ’Favor’, ’B’: ’Oppose’ Human distribution (U i , g): ’A’: 59.412, ’B’: 40.588 OpinionQA W26 Question (q i ): How safe, if at all, would you say your local community is from crime? Would you say it is Demographic group (g): ’RELIG’: ’Nothing in particular’ Answer Options (x k ): ’A’: ’Very safe’, ’B’: ’Somewhat safe’, ’C’: ’Not too safe’, ’D’: ’Not at all safe’, ’E’: ’Refused’ Human distribution (U i , g): ’A’: 25.795, ’B’: 55.701, ’C’: 14.579, ’D’: 3.925, ’E’: 0.0 SocSci210 nj5dx Question (q i ): You read an infographic highlighting the disadvantages faced by lower-class Americans, including statistics on income inequality, poverty rates, and disparities in health and life outcomes. After viewing it, you were asked: ́ The discrimination that people from lower classes have experienced is similar to that of other minority groups.Ónly return an integer from 1 to 7, nothing else, where 1 = "Strongly agree" and 7 = "Strongly disagree". Demographic group (g): ’AGE’: ’30-49’ Answer Options (x k ): ’1’: ’Strongly agree’, ’2’: ’Agree’, ’3’: ’Slightly Agree’, ’4’: ’Neutral’, ’5’: ’Slightly Disagree’, ’6’: ’Disagree’, ’7’: ’Strongly disagree’ Human distribution (U i , g): ’1’: 9.172, ’2’: 23.077, ’3’: 31.065, ’4’: 15.680, ’5’: 8.284, ’6’: 10.355, ’7’: 2.367 Figure 13: Test survey instance examples Prompt for Persona-conditioned Inference Role Definition (System Prompt): You are an individual with these characteristics: bio Prompt: **Question**: Question Estimate what percentage you would assign to each option. Follow these rules: 1. Use whole numbers from 0 to 100 2. Ensure the percentages sum to exactly 100 3. Only include the numbers (no % symbols) 4. Use this exact valid JSON format: json_format_str and do NOT include anything else. 5. Only output your final answer and nothing else. No explanations or intermediate steps are needed. Replace X with your estimated percentages for each option. **Answer**: Figure 14: Prompt template used for Persona-conditioned Simulation. Confidence LevelGSSOpinionQA W26OpinionQA W29 High0.8120.7290.842 Medium0.3200.4340.394 Low0.3930.3990.382 Table 7: Demographic feature extraction accuracy across different persona source datasets and confidence levels. FeatureGSSOpinionQA W26OpinionQA W29 Overall0.5530.3950.456 AGE0.4670.3000.422 EDUCATION0.3400.2670.310 INCOME0.2600.2460.347 MARITAL0.7310.4850.472 RACE0.7500.7830.871 CREGION0.3250.3420.250 RELIG0.3080.3930.403 SEX0.5830.8331.000 RELIGATTEND0.6940.2020.287 POLPARTY0.6420.5440.711 POLIDEOLOGY1.0000.3260.583 Table 8: Demographic feature extraction accuracy across different persona source datasets and demo- graphic keys.

Instant research discovery