Andrii Bidochko

• Updated: June 16, 2026
• 7 min read

Using Zero-Shot LLM-Generated Survey Data for Geographically Explicit Population Synthesis

Direct Answer

The paper demonstrates that zero‑shot large language models (LLMs) such as GPT‑4.1 and Gemini‑2.5‑Pro can generate synthetic health‑survey records that feed directly into an iterative proportional fitting (IPF) pipeline, producing geographically explicit synthetic populations. This matters because it opens a pathway to augment scarce survey data with AI‑generated inputs, potentially accelerating demographic modeling for policy, planning, and commercial analytics.

Background: Why This Problem Is Hard

Synthetic population generation sits at the intersection of demography, transportation planning, public‑health forecasting, and market analysis. Traditional pipelines rely on large, high‑quality household surveys (e.g., the Behavioral Risk Factor Surveillance System, BRFSS) that are expensive, infrequent, and often lack fine‑grained geographic coverage. When planners need tract‑level estimates for variables like smoking prevalence or diabetes rates, they must:

• Collect enough respondents in each small area to achieve statistical reliability.
• Reconcile disparate data sources (census, administrative records, surveys) through complex weighting schemes.
• Maintain privacy while preserving the statistical fidelity of the synthetic agents.

Existing approaches mitigate these issues by borrowing strength across regions using hierarchical Bayesian models or by augmenting surveys with auxiliary data (e.g., satellite imagery). However, these methods still depend on a baseline of real survey responses and often require bespoke statistical expertise. The bottleneck, therefore, is the scarcity of up‑to‑date, location‑specific micro‑data that can be fed into population synthesis engines.

What the Researchers Propose

The authors introduce a two‑stage framework that treats a zero‑shot LLM as a “virtual survey respondent generator.” In the first stage, the LLM receives a prompt describing the target state (Colorado or Mississippi) and the BRFSS questionnaire, then produces a set of synthetic individual records without any fine‑tuning or few‑shot examples. In the second stage, these AI‑generated records become the input micro‑data for a conventional IPF workflow, which aligns the synthetic sample with known marginal totals from the U.S. Census at the census‑tract level.

Key components of the framework include:

• Prompt Engine: Crafts state‑specific prompts that embed demographic priors (e.g., population size, age distribution) and the full BRFSS questionnaire.
• LLM Generator: Either GPT‑4.1 or Gemini‑2.5‑Pro produces a tabular dataset where each row mimics a survey respondent.
• IPF Synthesizer: Iteratively adjusts cell weights so that the synthetic sample matches marginal constraints (age, sex, race, income) for each tract.
• Validation Layer: Compares the resulting synthetic population against external benchmarks such as the American Community Survey (ACS) and state health department statistics.

How It Works in Practice

The operational workflow can be visualized as a linear pipeline:

1. Data Acquisition: Retrieve the 2023 BRFSS questionnaire and state‑level marginal tables from the Census.
2. Prompt Construction: For each target state, embed the marginal totals into a natural‑language prompt that instructs the LLM to “create 10,000 synthetic respondents reflecting Colorado’s demographics.”
3. Zero‑Shot Generation: Send the prompt to the LLM API. The model returns a CSV‑style list of respondents, each with values for health variables (e.g., smoking status, physical activity) and demographic attributes.
4. Pre‑Processing: Clean the AI‑generated table (remove impossible values, enforce categorical coding) to ensure compatibility with the IPF algorithm.
5. Iterative Proportional Fitting: Run the IPF routine, which repeatedly scales the synthetic table so that its marginal distributions align with the official tract‑level counts.
6. Post‑Processing & Validation: Aggregate the final synthetic population to the tract level and compute error metrics against known benchmarks (e.g., mean absolute error, spatial autocorrelation).

What distinguishes this approach from prior work is the elimination of a manual data‑collection step. The LLM acts as a “virtual field crew,” producing plausible micro‑data on demand, while the IPF stage remains unchanged—leveraging decades‑old, well‑understood statistical machinery.

{{IMAGE_PLACEHOLDER}}

Evaluation & Results

The researchers conducted experiments for two states with contrasting demographic profiles: Colorado (higher education levels, lower obesity rates) and Mississippi (higher poverty, higher smoking prevalence). They generated 20,000 synthetic respondents per state using both GPT‑4.1 and Gemini‑2.5‑Pro, then ran the IPF pipeline to produce tract‑level synthetic populations.

Key findings include:

• State‑Level Contrasts Captured: Both LLMs reproduced the broad north‑south health gradients (e.g., higher smoking rates in Mississippi) without any explicit geographic conditioning beyond the prompt.
• Variable‑Specific Performance Varies: Variables directly tied to census demographics (age, sex, race) aligned closely with ground truth, while more nuanced health behaviors (e.g., binge drinking) showed larger deviations.
• IPF Amplification Effects: In some cases, IPF corrected LLM biases (e.g., under‑representation of seniors), but in others it amplified errors, especially when the synthetic sample lacked sufficient variance for a given variable.
• Spatial Fidelity: At the census‑tract level, the synthetic populations reproduced spatial patterns of chronic disease prevalence reasonably well, as measured by Moran’s I and visual heat‑maps.

Overall, the experiments suggest that zero‑shot LLM‑generated survey data can serve as a supplementary input for population synthesis, but it is not yet a full replacement for real survey data.

Why This Matters for AI Systems and Agents

For AI practitioners building agents that rely on realistic demographic simulations—whether for urban mobility, public‑health forecasting, or targeted marketing—the ability to generate synthetic micro‑data on demand is a game‑changer. It reduces dependence on costly field surveys and enables rapid scenario testing (e.g., “What if a new health policy reduces smoking by 5% in rural tracts?”). Moreover, the workflow integrates seamlessly with existing agent‑orchestration platforms that already support IPF or similar weighting engines.

Practically, teams can embed the LLM‑prompt step into a Workflow automation studio to trigger synthetic data generation whenever new marginal data become available. The resulting synthetic agents can then be fed into Enterprise AI platforms for downstream analytics, simulation, or decision support.

In the context of conversational AI, a virtual assistant could query the synthetic population to answer “What is the estimated prevalence of diabetes among 30‑40‑year‑olds in Denver?” without exposing any real personal data, thereby preserving privacy while delivering actionable insights.

What Comes Next

While promising, the study highlights several open challenges:

• Bias Propagation: LLMs inherit biases from their training corpora, which can manifest as systematic over‑ or under‑estimation of certain health outcomes.
• Variable Coverage: Not all survey items are equally well modeled; rare or highly contextual questions may require few‑shot prompting or fine‑tuning.
• Scalability: Generating millions of records for a national‑scale synthesis will demand efficient batching and cost‑management strategies.
• Validation Frameworks: Robust, automated validation pipelines are needed to flag when LLM‑generated data diverge dangerously from known benchmarks.

Future research could explore hybrid pipelines that combine a small seed of real survey responses with LLM‑augmented synthetic records, or employ reinforcement learning to iteratively improve prompt quality based on downstream IPF error signals.

Potential applications extend beyond public health. For example, AI marketing agents could use synthetic demographic profiles to simulate campaign reach in under‑sampled neighborhoods, while Openclaw bots could test policy interventions in a sandbox environment before real‑world rollout.

Developers interested in experimenting with this approach can start by integrating the LLM generation step with the Chroma DB integration for efficient storage and retrieval of synthetic records, then connect to the Telegram integration on UBOS to receive real‑time synthesis updates.

For a deeper dive into the methodology and raw results, consult the original arXiv paper.

Andrii Bidochko

CTO UBOS

Andrii Bidochko is an AI entrepreneur and researcher focused on AI agents, reinforcement learning, and autonomous systems. He writes about the technologies shaping the future of machine intelligence, from frontier models and agent architectures to real-world AI applications.