Generative Synthetic Populations (GSP)

• Generative Synthetic Populations are techniques that create detailed synthetic data sets representing individuals, households, and spatial ecosystems by reconciling microdata with aggregate constraints for simulation.
• They employ diverse methods including maximum-entropy, copula-based transfer, combinatorial optimization, and deep generative models to balance realism, diversity, and exact fit.
• These methods enable rigorous microsimulation, agent-based modeling, and policy testing by generating synthetic populations that reflect complex demographic and spatial structures.

Searching arXiv for recent and foundational papers on generative synthetic populations to ground the encyclopedia article. First, I’ll look for papers explicitly using the term “synthetic population” and related generative methods. Generative Synthetic Populations (GSP) denotes a family of methods for constructing synthetic individuals, households, linked entities, or broader synthetic ecosystems for microsimulation and agent-based modeling, typically from census marginals, survey microdata, or both. In the literature, the generated object ranges from categorical person records and household rosters to spatial ecosystems and LLM-driven synthetic personas; correspondingly, GSP spans aggregate-only maximum-entropy and constraint-programming methods, microdata-plus-calibration pipelines, deep generative tabular models, and support-coverage-oriented persona generation (Mahmood et al., 2024, Gallagher et al., 2017, Esposti, 1 Jul 2026, Paglieri et al., 3 Feb 2026).

1. Scope and conceptual boundaries

Within agent-based simulation, synthetic populations are used to represent “the structure, behaviour, and interactions of individuals,” while synthetic ecosystems extend that object to include spatial placement and environmental components such as schools and workplaces (Mahmood et al., 2024, Gallagher et al., 2017). In this literature, a synthetic population is not merely anonymized tabular data. It is a simulation substrate: a population of agents, often geographically constrained, with demographic, household, behavioral, or institutional attributes suitable for downstream modeling (Wu et al., 2016, Neekhra et al., 2023).

A recurrent distinction concerns the target of generation. Much of the literature treats GSP as a density-matching problem: reproduce marginals, cross-tabulations, or other aggregate constraints of a reference population (Mahmood et al., 2024, Jutras-Dubé et al., 2023). Other work argues that some GSP settings are instead support-coverage problems, where the objective is to span the space of plausible personas, including rare combinations and long-tail behaviors, rather than replicate the most probable cases (Paglieri et al., 3 Feb 2026). This suggests that “population realism” is not a single criterion; it depends on whether the synthetic population is intended for calibration, exploration, stress-testing, or scenario generation.

The term also has a broader contemporary meaning. One line of work defines GSP as the convergence of synthetic population synthesis, agent-based modelling, and LLM agents, and proposes that an LLM-driven synthetic population should first be shown to be controllable before any claim about external validity is attempted (Esposti, 1 Jul 2026). In that formulation, a GSP is both a population model and an instrument: it should respond to stimuli in an ordered, replicable, and group-structured way (Esposti, 1 Jul 2026).

2. Data regimes, synthesis targets, and structural units

A central organizing axis in GSP is the available data regime. Several methods are explicitly sample-free or aggregate-driven: they synthesize micro-level populations from marginals, contingency tables, or other published aggregates without using target-area microdata as the primary generative basis (Mahmood et al., 2024, Esposti, 28 Mar 2026, Petit et al., 8 Dec 2025). Other methods are hybrid. They learn dependence structure from microdata while imposing target marginals from aggregate sources, as in copula-based transfer from one geography to another or WGAN-based localization from EU-SILC microdata to city populations (Jutras-Dubé et al., 2023, Falck, 27 Jan 2025). A further class combines multiple surveys or administrative sources with partially overlapping schemas, using conditional generation to enrich a base population with additional variables (Neekhra et al., 2023, Wan et al., 2019).

The target representation also varies substantially. Some frameworks generate only individual-level records over categorical attributes, such as sex, age-bin, race, income-bin, occupation, or education (Wu et al., 2016). Others explicitly generate households containing persons, with a mapping from person records to household entities (Mahmood et al., 2024, Qian et al., 2024). Linked-entity formulations generalize this further to bipartite populations such as dwellings–households, households–cars, or workers–firms, where the synthetic population is a triplet $\langle \hat{A}, \hat{B}, L \rangle$ rather than a single table (Thiriot et al., 2020). At the largest scope, synthetic ecosystems include households, individuals, locations, and environmental components, and SPEW has generated over five billion human agents across approximately 100,000 geographic regions in about 70 countries (Gallagher et al., 2017).

Spatial granularity ranges from census tracts and PUMAs to districts, states, and country-scale populations. Delaware tract synthesis from state-level PUMS and ACS marginals exemplifies small-area joint household–individual generation (Qian et al., 2024). India-focused systems target districts and finer administrative units within districts (Neekhra et al., 2022, Neekhra et al., 2023). New York State synthesis at nearly 20 million individuals and 7.5 million households illustrates state-scale generation of multi-person households (Yang et al., 13 Aug 2025). This suggests that GSP is less a single algorithmic family than a design space defined by data availability, structural units, and intended simulation scale.

3. Statistical, optimization, and declarative formulations

One foundational line formulates GSP as maximum-entropy inference from aggregate constraints. In the aggregate-only setting, the target is a distribution $p(x)$ over the categorical tuple space $X = D_1 \times \cdots \times D_K$ such that $\mathbb{E}_{p}[f_j] = \alpha_j$ for unary, binary, or ternary indicator constraints (Esposti, 28 Mar 2026). The corresponding exponential-family model is

$$p_\lambda(x)=\frac{1}{Z(\lambda)}\exp\!\left(\sum_{j=1}^{m}\lambda_j f_j(x)\right),$$

with dual objective $\log Z(\lambda)-\alpha^\top\lambda$ (Esposti, 28 Mar 2026). Earlier work likewise cast synthetic population generation as a maximum-entropy problem over categorical tuples, using selected pattern frequencies as sufficient statistics and tuple-block structure for tractable inference (Wu et al., 2016). Recent work replaces exact expectation computation by persistent Gibbs sampling, yielding a scalable sample-free MaxEnt solver whose runtime scales as $O(K)$ rather than $O(|\mathcal{X}|)$, with $\mathrm{MRE}=0.03$ on Syn-ISTAT and an $86.8\times$ effective sample size advantage over generalized raking (Esposti, 28 Mar 2026).

A second line separates marginals from dependence using copulas. The transferable-copula formulation assumes that source and target populations share a copula $p(x)$0, even if their marginals differ, so that

$p(x)$1

can be synthesized by learning dependence in copula-normalized space from one geography and injecting target marginals by inverse CDF mapping (Jutras-Dubé et al., 2023). SynC extends this logic into a multi-stage framework that combines Gaussian copula sampling, predictive models for merging variable batches, and post hoc scaling to match marginal constraints (Wan et al., 2019). These approaches are particularly suited to the frequent GSP setting in which only empirical marginal distributions are available for the target area, while microdata exist only for another, related population (Jutras-Dubé et al., 2023).

A third line treats GSP as combinatorial or declarative optimization. Large-scale hierarchical population synthesis can be written as a multi-objective search over entire candidate populations $p(x)$2, minimizing per-attribute discrepancies

$p(x)$3

with NSGA-II, while preserving a two-layer hierarchy of persons nested within households (Mahmood et al., 2024). Constraint-programming approaches instead encode aggregate distributions and logical admissibility directly over person-level variables $p(x)$4, with exact integer target counts obtained by largest-remainder rounding and deviations minimized under hard individual constraints (Petit et al., 8 Dec 2025). The linked-entity problem has also been formalized analytically: if $p(x)$5, $p(x)$6, degree distributions, and pairing probabilities $p(x)$7 are all supplied, the problem is typically over-constrained, and consistency requires relaxation of some inputs before direct generation of the linked bipartite population (Thiriot et al., 2020).

These formulations differ sharply in what they preserve exactly. Maximum-entropy methods privilege moment matching under minimal additional assumptions (Wu et al., 2016, Esposti, 28 Mar 2026). Copula methods privilege target marginals while transferring dependence structure (Jutras-Dubé et al., 2023, Wan et al., 2019). Multi-objective search keeps multiple fitting targets separate during evolution and selects a final solution from the Pareto set by a weighted preference rule (Mahmood et al., 2024). Constraint programming prioritizes logical validity and exact control over declared counts when feasible (Petit et al., 8 Dec 2025). This suggests that “realism” in GSP is model-dependent: it may mean entropy-maximal consistency with published tables, transferable dependence with corrected marginals, Pareto-optimal joint fit, or exact satisfaction of declarative constraints.

4. Deep generative, multi-source, and semantic-conditioned models

Deep generative work in GSP has concentrated on VAEs, GANs, WGANs, CTGAN-style models, and specialized household generators. One major theme is that deep models can recover valid but unobserved combinations—sampling zeros—but may also generate infeasible combinations—structural zeros (Kim et al., 2022). To address this, feasibility-aware regularizers have been added to VAEs and WGAN-GP, including boundary distance regularization

$p(x)$8

and average distance regularization

$p(x)$9

which trade off feasibility against diversity (Kim et al., 2022). In the South Korea experiment, the best regularized VAE generated an additional $X = D_1 \times \cdots \times D_K$0 of the population ignored by the sample with $X = D_1 \times \cdots \times D_K$1 precision, while the best regularized WGAN generated $X = D_1 \times \cdots \times D_K$2 of the ignored population with $X = D_1 \times \cdots \times D_K$3 precision (Kim et al., 2022).

Another theme is explicit household modeling. A deep generative framework for joint households and individuals reshapes each household and all of its members into one fixed-width row, trains a VAE on that representation, and then fine-tunes latent codes so that decoded households match tract-level marginals (Qian et al., 2024). Its decoupled reconstruction objective replaces one-to-one row reconstruction by resemblance to any plausible household in the training support, via a differentiable soft nearest-neighbor BCE scheme (Qian et al., 2024). Relatedly, ciDATGAN-based synthesis for New York State trains a separate household-size-specific model on flattened multi-person household records, conditions generation on residence area, age, and race, and reports nearly 20 million individuals and 7.5 million households, with entropy-based diversity gains of $X = D_1 \times \cdots \times D_K$4 over PUMS and $X = D_1 \times \cdots \times D_K$5 over a Popgen benchmark (Yang et al., 13 Aug 2025).

Deep generative work has also moved toward spatial and multi-source settings. A WGAN-GP workflow trained on EU-SILC microdata is used to generate spatial synthetic populations for Helsinki and Thessaloniki, with balancing either by weights or by external local aggregate statistics (Falck, 27 Jan 2025). A joint multi-source WGAN with two critics and an inverse gradient penalty regularizer improves both recall and precision relative to a sequential baseline, with recall increasing by $X = D_1 \times \cdots \times D_K$6, precision by $X = D_1 \times \cdots \times D_K$7, and a final five-metric similarity score of $X = D_1 \times \cdots \times D_K$8 compared to $X = D_1 \times \cdots \times D_K$9 for the sequential method (Abbasi et al., 17 Feb 2026). These results suggest that simultaneous integration and synthesis can preserve cross-source interactions better than a fuse-then-generate pipeline (Abbasi et al., 17 Feb 2026).

A distinct frontier introduces semantic conditioning. SemaPop generates a persona text $\mathbb{E}_{p}[f_j] = \alpha_j$0 for each survey record, embeds it as $\mathbb{E}_{p}[f_j] = \alpha_j$1, adapts it into a condition vector, and injects that semantics into a WGAN-GP via FiLM modulation in the generator and projection conditioning in the critic (Qin et al., 12 Feb 2026). More radically, Persona Generators optimize support coverage rather than density matching. They define

$\mathbb{E}_{p}[f_j] = \alpha_j$2

and optimize $\mathbb{E}_{p}[f_j] = \alpha_j$3 to maximize diversity metrics on simulated questionnaire-response embeddings rather than match an observed target joint distribution (Paglieri et al., 3 Feb 2026). Finally, LLM-driven synthetic populations have been treated as instruments whose controllability must be calibrated through repeated stimuli, placebo conditions, and temperature sweeps before any substantive external-validity claim is made (Esposti, 1 Jul 2026). Taken together, these papers indicate an expansion of GSP from tabular demographic synthesis toward semantically interpretable and behaviorally conditioned synthetic populations.

5. Evaluation criteria, diagnostics, and recurrent trade-offs

Evaluation in GSP is heterogeneous because different methods optimize different objects. Classical distributional fidelity is often measured with SRMSE, RMSE, $\mathbb{E}_{p}[f_j] = \alpha_j$4, Pearson correlation, KL divergence, Jensen–Shannon distance, chi-square tests, or KS tests (Jutras-Dubé et al., 2023, Falck, 27 Jan 2025, Qian et al., 2024, Neekhra et al., 2022). For transfer learning across geographies, SRMSE is frequently computed on $\mathbb{E}_{p}[f_j] = \alpha_j$5-way marginals, with $\mathbb{E}_{p}[f_j] = \alpha_j$6, to separate marginal fit from higher-order dependency preservation (Jutras-Dubé et al., 2023). In aggregate-only MaxEnt work, the principal training metric is mean relative error,

$\mathbb{E}_{p}[f_j] = \alpha_j$7

while diversity is assessed with effective sample size

$\mathbb{E}_{p}[f_j] = \alpha_j$8

entropy, and number of unique profiles (Esposti, 28 Mar 2026).

Support-based evaluation has become increasingly important. In feasibility-aware deep synthesis, precision is the proportion of generated combinations that exist in the real population support, and recall is the proportion of real population combinations covered by the generated support; structural-zero rate is therefore $\mathbb{E}_{p}[f_j] = \alpha_j$9 (Kim et al., 2022). Copula-transfer work complements SRMSE with sampled zeros, defined as realistic combinations absent from the training sample but present in the reference population, to test whether the synthesizer can go beyond simple replication (Jutras-Dubé et al., 2023). Diversity can also be quantified through entropy over household-member combinations, as in the New York household synthesis framework (Yang et al., 13 Aug 2025), or through support-coverage metrics such as convex hull volume, coverage, minimum and average pairwise distance, dispersion, and KL divergence to a Sobol quasi-random reference in support-covering persona generation (Paglieri et al., 3 Feb 2026).

Several papers make explicit that low marginal error is not sufficient. In the Helsinki/Thessaloniki WGAN study, Pearson, $$p_\lambda(x)=\frac{1}{Z(\lambda)}\exp\!\left(\sum_{j=1}^{m}\lambda_j f_j(x)\right),$$0, and SRMSE are described as “shallow,” while Bland–Altman plots are used to reveal hidden over- and under-representation in high-dimensional outputs (Falck, 27 Jan 2025). In the LLM-driven calibration study, validity is assessed not by human comparison but by seven pre-registered internal criteria: persona fidelity, cross-replica stability, noise floor, replica bias, placebo specificity, sensitivity to stimulus valence, ordering under known condition structure, and receipt check (Esposti, 1 Jul 2026). This suggests that evaluation in GSP depends on whether the target failure mode is marginal mismatch, loss of support, infeasible records, poor diversity, hidden multivariate distortion, or unstable behavioral response.

A persistent trade-off cuts across the literature: exact fit versus diversity. Reweighting and raking can match training marginals exactly yet collapse effective diversity, with $$p_\lambda(x)=\frac{1}{Z(\lambda)}\exp\!\left(\sum_{j=1}^{m}\lambda_j f_j(x)\right),$$1 dropping to $$p_\lambda(x)=\frac{1}{Z(\lambda)}\exp\!\left(\sum_{j=1}^{m}\lambda_j f_j(x)\right),$$2 on Syn-ISTAT, while a generative MaxEnt sampler retains $$p_\lambda(x)=\frac{1}{Z(\lambda)}\exp\!\left(\sum_{j=1}^{m}\lambda_j f_j(x)\right),$$3 at the cost of approximate rather than exact training fit (Esposti, 28 Mar 2026). Likewise, increasing diversity in deep generators can increase structural zeros, and improving local aggregate calibration can worsen fringe-profile distortion (Kim et al., 2022, Falck, 27 Jan 2025). A plausible implication is that GSP evaluation should be multi-objective by design rather than reduced to a single goodness-of-fit score.

6. Applications, limitations, and emerging directions

Applications are broad and often simulation-driven. Household-based ABMs, transport models, public-health simulation, infectious-disease modeling, and urban digital twins are recurrent targets (Mahmood et al., 2024, Neekhra et al., 2023, Gallagher et al., 2017). Country-scale synthesis has been demonstrated for India by combining census marginals, survey microdata, and geospatial grids to generate family-structured, geolocated populations for BharatSim-style infectious-disease simulation (Neekhra et al., 2023, Neekhra et al., 2022). Spatial WGAN pipelines are motivated by urban planning, public health, and economic forecasting (Falck, 27 Jan 2025). Constraint-programmed aggregate-only populations are deployed as queryable digital twins for polling and territorial economic intelligence (Petit et al., 8 Dec 2025). Support-covering persona generation is aimed at AI evaluation, red-teaming, novel-product testing, and speculative or future scenarios where a faithful empirical target distribution may not exist (Paglieri et al., 3 Feb 2026).

The limitations are equally consistent. Household realism is often partial. NSGA-II-based hierarchical synthesis explicitly notes that household allocation is not yet sensitive to ethnicity, religion, or other within-household dependencies (Mahmood et al., 2024). Spatial GAN work operates at regional or city-targeting level rather than fine-grained residential allocation and does not model households structurally (Falck, 27 Jan 2025). Deep household generators can still produce record-level contradictions; one VAE-based tract synthesizer reports 158 households, or $$p_\lambda(x)=\frac{1}{Z(\lambda)}\exp\!\left(\sum_{j=1}^{m}\lambda_j f_j(x)\right),$$4, with inconsistent $$p_\lambda(x)=\frac{1}{Z(\lambda)}\exp\!\left(\sum_{j=1}^{m}\lambda_j f_j(x)\right),$$5 and member-age relationships in one Delaware tract (Qian et al., 2024). Sample-free declarative methods provide logical guarantees for encoded constraints, but only with respect to the constraints actually specified, and they do not yet model richer relational structures such as households and firms jointly (Petit et al., 8 Dec 2025).

Fairness, rare groups, and privacy remain contested areas. The EU-SILC WGAN study explicitly warns that deep generative methods may under-represent fringe profiles and that such suppression can lead to discrimination in agent-based simulations; it finds over-representation of the most frequent self-perceived-health category and under-representation of less frequent values under a WGAN-impute pipeline (Falck, 27 Jan 2025). Feasibility-aware GAN/VAE work frames the central problem as balancing sampling zeros against structural zeros rather than maximizing one at the expense of the other (Kim et al., 2022). Privacy benefits are often asserted from the synthetic nature of the data, but formal disclosure-risk analysis, membership inference testing, or differential privacy guarantees are usually absent (Falck, 27 Jan 2025, Wu et al., 2016).

Several emerging directions are explicit in the literature. Hierarchical population synthesis is being extended toward richer within-household dependence and parallel processing (Mahmood et al., 2024). Spatial generative work points toward better preservation of rare profiles, stronger target-area constraints, and possibly diffusion-style or neural-manifold methods (Falck, 27 Jan 2025). Semantic-persona conditioning and support-coverage optimization indicate a shift toward controllable synthetic populations whose latent behavioral structure can be edited or stress-tested (Qin et al., 12 Feb 2026, Paglieri et al., 3 Feb 2026). LLM-driven GSP adds a further requirement: before such systems are used as substitutes for human populations, they should be calibrated as instruments with known response functions and quantified noise floors (Esposti, 1 Jul 2026).

Taken together, these papers suggest that GSP is evolving along three partially independent axes. The first is statistical: better matching of marginals, joints, and transferable dependence. The second is structural: richer households, linked entities, and spatial ecosystems. The third is semantic: explicit modeling of latent behavioral patterns, support coverage, and controllability. Contemporary GSP research increasingly treats these not as interchangeable goals but as distinct design choices to be balanced according to the intended use of the synthetic population.