Synthetic population data, that are known as pseudo data

of household members, are widely used in agent-based simulations for

urban planning and policy analysis. However, the stability of synthetic

population data generation across multiple runs has not been sufficiently

investigated. This study proposes a framework for evaluating the stability of synthetic population data generation using multiple independently

generated datasets. The analysis focuses on three household attributes:

the industry category of the household head, the maximum age within the

household, and the age difference between spouses. The results demonstrate that the stability of synthetic population data generation depends

on attribute characteristics and the underlying distribution. In particular, attributes with dense distributions in small areas tend to produce

more stable outcomes, whereas attributes with sparse distributions are

more sensitive to stochastic variation. These findings indicate that stability cannot be adequately evaluated using a single evaluation setting.

Furthermore, the agreement rates based on modal attribute values should

be interpreted as relative measures rather than absolute indicators, particularly when comparing stability across different attributes.

We propose a nationwide workplace assignment framework that simultaneously considers the national statistics of workers in resi- dence municipalities, workplace municipalities, and industry of workers. We address the lack of workplace-side consistency in Japanese synthetic populations that were mainly synthesized from residence-side statistics. We employ a multidimensional iterative proportional fitting (IPF) proce- dure to estimate a consistent joint distribution while keeping the residence- workplace-industry statistics that are publicly released only for large mu- nicipalities. Workers are then assigned workplace municipalities stochas- tically conditional on residence and industry. Workplace locations are further refined to intra-municipality small areas using Economic Census. Experiments show improved agreement with official statistics at both municipality and small-area levels, enabling realistic nationwide daytime population simulations.

Synthetic cities are increasingly used in simulation, scenario analysis, and policy evaluation. Existing generation methods can reproduce selected structural properties of urban systems, but many rely on simplified road representations and provide limited support for combining calibration to a reference city with high-level design requirements. This paper presents a framework for generating directed and attributed synthetic transport networks using large language models. The proposed method combines natural-language requirements with quantitative validation against a reference structure and produces road networks in which directionality, road classes, and edge weights are represented explicitly. The aim is to generate synthetic urban transport structures that preserve selected properties of a reference area while being better suited to routing, accessibility analysis, and agent-based simulation.