A Text-Based City for AI Agents: The Urban Generative Intelligence Proposal

Architecture of the foundational platform, from the preprint. CC BY 4.0.

Imagine giving a language model a city to live in. A preprint on arXiv proposes exactly that: Urban Generative Intelligence (UGI), a platform where AI agents run on a large language model act inside a realistic, text-described simulation of a city, learn from the feedback, and are used on tasks from planning transportation systems to assisting policy-making. The paper is a proposal and a framework; as published, it describes the design, an evaluation framework and design cases, and says its code with demonstrations would be released soon at a GitHub repository.

The problem

Cities are complex systems of physical, social, economic and environmental layers, and rapid urbanization brings congestion, pollution, resource scarcity, strain on infrastructure, inequality and housing problems, with climate change adding more. The paper traces a sequence of tools: big data, artificial intelligence, urban computing, digital twins and city simulations. Each stage builds on the last, yet there is still a gap between what the tools can do and solving cities' systemic problems. The authors see large language models (LLMs) as a chance to close it, because they can reason about context and find patterns in urban discourse.

Three layers

1. Open digital infrastructure. It collects spatial and temporal data on points of interest, roads and subways, individual and collective movement and traffic. A city simulator called Mirage models how people, things and places interact and can run hypothetical scenarios, and a knowledge graph, UrbanKG, joins the data streams and stores facts such as which place borders another.
2. A language interface. A standard natural-language protocol lets models and agents use the simulator and the knowledge graph without custom code, so developers can plug in language models more easily.
3. Generative intelligence. On top sits CityGPT, a model built for urban problems, and agents for mobility, economy, community and society.

How CityGPT is built

The authors argue that general open models do poorly on urban tasks for lack of domain knowledge. They first continue pre-training a general model on general text plus urban material, including text generated from knowledge-graph facts, geographic data in GeoJSON format and records of human behavior, then apply supervised fine-tuning and a preference method (DPO) so it follows instructions and returns structured output such as JSON. As an example, they continued pre-training Baichuan2-7B on one machine with eight A100 GPUs for about a week.

Training procedure of CityGPT, from the preprint. CC BY 4.0.

Judging urban intelligence

The proposed evaluation has three levels, from knowledge to action:

• Level 1: automated multiple-choice questions on urban knowledge, drawn from general exams and generated from the training corpus.
• Level 2: simple reasoning on new questions written by volunteer experts, such as commonsense and route planning.
• Level 3: planning and decision-making by city agents on four tasks: next-location prediction, map-free navigation between points of interest, daily schedule planning and society simulation, each with hundreds of samples scored by success rate.

The proposed evaluation framework, from the preprint. CC BY 4.0.

What it could be used for

• Transport: travel surveys are costly and infrequent, and rule-based mobility generators lack an understanding of intent, so agents could produce more realistic, intention-aware travel behavior.
• Business: combining the knowledge graph with a city model for site selection.
• Economy: agent-based simulations of macroeconomic behavior and policy testing.
• Society: simulating individuals and populations, testing interventions and preparing for emergencies for which data are scarce.

Open challenges

The authors list scaling to a large city, where simulating many language-model agents is computationally costly (techniques such as batch prompting are cited as promising), keeping the environment open so users can reshape a real city or build a new one, and building a developer community across data science, simulation, language models and urban planning.

Sources

• "Urban Generative Intelligence (UGI): A Foundational Platform for Agents in Embodied City Environment", arXiv preprint (https://arxiv.org/abs/2312.11813), licensed CC BY 4.0; the authors are not named in the text provided. Changed here: condensed and rewritten in hubnx's own words; references, most figures and the technical detail of the infrastructure left out. Figures are the preprint's own.
• Authors of the original: Fengli Xu, Jun Zhang, Chen Gao.