Surpassing 90% of urban planners, Tsinghua University, MIT, etc. propose a new paradigm of human-machine collaboration.

Tsinghua University, in collaboration with institutions such as MIT, has proposed a new framework of "Large Language Model + Planner." AI accompanies the entire process like an assistant, from discussing requirements and drawing blueprints to rehearsing the effects with virtual residents, making urban planning faster and more scientific. Experiments show that AI can outperform over 90% of human planners. In the future, humans and machines will each exert their strengths to design cities that are more livable and equitable.

Facing the increasingly complex urban system and diverse social needs, traditional urban planning methods are encountering bottlenecks.

Today, artificial intelligence (AI) is bringing disruptive innovations to this ancient and important field.

Recently, an interdisciplinary team composed of scholars from the Research Center for Urban Science and Computing in the Department of Electronic Engineering at Tsinghua University, the School of Architecture at Tsinghua University, the Senseable City Lab at the Massachusetts Institute of Technology (MIT), and Northeastern University in the United States published a perspective article in the international leading journal Nature Computational Science. For the first time, they systematically proposed an intelligent urban planning framework driven by large language models (LLMs).

Paper link: https://www.nature.com/articles/s43588-025-00846-1

This framework deeply integrates the powerful computing, reasoning, and generation capabilities of AI with the professional experience and creativity of human planners. It aims to transform AI into a "smart planning assistant" for humans to jointly address the complex challenges in modern urban planning, and opens a new paradigm of human - machine collaboration for a more efficient, innovative, and responsive urban design process.

The first author of the paper is Zheng Yu, a doctoral student in the Department of Electronic Engineering at Tsinghua University. The corresponding authors are Professor Li Yong from the Department of Electronic Engineering at Tsinghua University, Assistant Professor Lin Yuming from the School of Architecture at Tsinghua University, and Associate Professor Qi R. Wang from the Department of Environmental Engineering at Northeastern University.

The collaborators include Assistant Professor Xu Fengli from the Department of Electronic Engineering at Tsinghua University, as well as Researcher Paolo Santi and Professor Carlo Ratti from the Senseable City Lab at MIT.

Evolution and Bottlenecks of Urban Planning

The theory and practice of urban planning have been continuously evolving. It has evolved from the early "artistic design" that focused on physical space and aesthetic form to the "scientific planning" after World War II, which regards urban planning as a complex system and uses scientific models for analysis.

However, these methods are facing new challenges today:

On the one hand, the planning process still centers around planners, and the breadth and depth of public participation are limited.

On the other hand, the evaluation of planning schemes is often qualitative, subjective, and lagging, making it difficult to make scientific quantitative decisions and rapid iterations.

In recent years, traditional AI models represented by generative adversarial networks (GANs) and reinforcement learning (RL) have begun to be applied in urban planning and have shown potential in generating street networks, functional zoning, etc.

However, these models are usually designed for specific tasks, with a narrow knowledge base, and are difficult to cope with the increasing interdisciplinary complexity of modern urban planning.

The emergence of large language models (LLMs) brings a historical opportunity to break through this bottleneck with their powerful knowledge integration, logical reasoning, and multimodal generation capabilities.

New Process of LLM - Driven Urban Planning

In response to the deficiencies of traditional methods, the research team innovatively proposed a closed - loop framework consisting of three core stages: Conceptualization, Generation, and Evaluation.

This framework is jointly driven by large language models, large vision models (VLMs), and large model agents (LLM Agents), providing full - process intelligent assistance for human planners.

Figure 1: The proposed LLM - driven urban planning framework

Conceptual Design: LLM Becomes a "Planning Advisor" with Interdisciplinary Knowledge

In the early stage of planning, planners input text information such as requirements, constraints, and guidelines.

The LLM, pre - trained on a large amount of data, can deeply integrate knowledge in multiple fields such as geography, society, and economy, and conduct multiple rounds of "dialogue" with planners.

It can not only propose innovative conceptual ideas but also reason based on complex contexts, generating detailed planning description texts and preliminary spatial concept sketches, greatly improving the efficiency and depth of the conceptual design stage.

Figure 2: Flowchart of LLM - based urban conceptual design

Scheme Generation: VLM Transforms into a "Visual Designer" to Convert Text into Blueprints

This framework uses large vision models (VLMs) to transform abstract text concepts into specific and visual urban design schemes.

Planners can precisely describe planning concepts and constraints through text prompts. The VLM fine - tuned with urban design data can generate detailed visual outputs, such as land use layouts, building outlines, and even realistic 3D urban scenes, while ensuring that the design complies with real - world constraints such as geography.

Figure 3: Schematic diagram of urban scheme generation

Effect Evaluation: LLM Agents Build a "Virtual City" to Rehearse Future Life

To scientifically evaluate planning schemes, the framework introduces LLM agents for urban dynamic simulation.

Researchers set different demographic characteristics (such as age and occupation) for the agents, allowing them to simulate the daily travel and facility use activities of residents in the generated virtual city.

By analyzing these simulated behaviors, multi - dimensional quantitative evaluation indicators such as traffic distance, facility utilization rate, carbon emissions, and social equity can be obtained, providing scientific and forward - looking feedback for the iterative optimization of planning schemes.

Figure 4: Urban planning effect evaluation scheme based on LLM & VLM agents

Initial Success: AI Shows Potential to Surpass Human Experts

To verify the feasibility of the core capabilities of this framework, the Research Center for Urban Science and Computing in the Department of Electronic Engineering at Tsinghua University has continuously released a series of large urban multimodal language - vision models such as CityGPT, CityBench, and UrbanLLaVA, as well as urban embodied simulation platforms and social simulation systems such as UrbanWord, EmbodiedCity, and AgentSociety, laying a technical foundation for urban planning and social governance in the era of large models.

The research team conducted a series of proof - of - concept experiments on urban planning in the LLM era.

In a test, researchers asked the LLM to answer questions from the professional qualification exam for urban planners. The results showed that the largest - scale LLM performed better than the top 10% of human planners in answering questions about complex planning concepts, proving its great potential in the conceptualization stage.

In the simulation test of the evaluation stage, the team used LLM agents to simulate the facility access behaviors of residents in two communities in New York and Chicago, USA.

The simulation results showed that the hot - spot areas visited by the agents were highly consistent with the real resident flow data, proving the accuracy and effectiveness of LLM agents in predicting the actual impact of planning schemes.

Figure 5: Schematic diagram of urban planning effects generated by LLM

Challenges and Prospects: Building a Future City with Human - Machine Collaboration

The research team finally emphasized that this framework is not intended to replace human planners but aims to establish a new workflow of human - machine collaboration.

In this model, planners can be liberated from tedious data processing and drawing work and focus more on innovation, ethical considerations, and communication with various stakeholders, while AI is responsible for efficiently completing concept integration, scheme generation, and simulation evaluation.

At the same time, the article also pointed out the challenges faced by this technical route, including the scarcity of high - quality urban design data, the huge demand for computing resources, and potential geographical and social biases in the models.

Future research needs to establish open data platforms, develop more efficient dedicated models, and design fairness algorithms to ensure that AI technology can serve all urban environments fairly and inclusively.

We can expect that in the near future, urban planners, with the help of powerful AI assistants, will be able to design more efficient, livable, and sustainable cities faster and better, fully unleashing human creativity to shape our shared urban home.

References

https://www.nature.com/articles/s43588-025-00846-1

This article is from the WeChat official account "New Intelligence Yuan". Author: LRST. Republished by 36Kr with permission.