Loading 3D visualization...

GenWorld

LLM-Ready Urban Simulation Platform

Empirically-grounded simulation with synthetic populations validated against census, mobility, and time-use survey data

196,608

Individuals

90,104

Households

0.00%

Pop. Error

132K+

Trajectories

Paper Code Explore 3D Map Cite

Abstract

We present GenWorld, an LLM-ready urban simulation platform with empirically-grounded synthetic populations. The platform generates 196,608 individuals across 90,104 households with census-validated demographics (0.00% population error), building-level location assignments, and time-use survey grounded activity schedules. Our validation against Japanese Census data, YJMob100K mobility dataset, and e-Stat time-use surveys demonstrates high fidelity in demographic distributions, commuting patterns, and temporal activity patterns.

5-Step Data Pipeline

1

Population Synthesis

IPF algorithm with census constraints

2

POI Classification

17 categories from OSM data

3

Location Assignment

Home → School → Work

4

LLM Distillation

Offline behavioral priors

5

Activity Simulation

Time-use survey grounded

Validation Results

0.00%

Population Error

Census-validated demographic distribution

Population Validation
0.13%

Household Error

Household size distribution validation

Commute Distance
0.9975

Age Correlation

Age distribution accuracy

Activity Validation

Additional Visualizations

24-Hour Activity Pattern
24-hour activity radial plot showing temporal patterns across the day
Spatial-Temporal Heatmap
Spatial-temporal activity distribution heatmap
Platform Overview
GenWorld platform architecture and data flow

Citation

@article{genworld2026,
  title={GenWorld: An LLM-Ready Urban Simulation Platform with
         Empirically-Grounded Synthetic Populations},
  author={Li, Gen and Lan, Jieyuan and Xu, Pengcheng and
          Wu, Zongyuan and Ogura, Masaki and Feng, Tao},
  journal={arXiv preprint},
  year={2026}
}

Hiroshima University Graduate School of Advanced Science and Engineering