Urban Mobility Forecasting Reimagined: Can LLM-Based Agents Transform Transportation Planning?

Urban mobility systems are, at their core, reflections of human behavior. Every trip, whether it is a daily commute, a leisure outing, or a spontaneous errand, is shaped by complex, context-dependent decisions. For decades, transportation planning has relied on models that attempt to simplify and predict these behaviors. However, capturing the true nature of human decision-making remains one of the most persistent challenges in the field.

Traditional approaches, such as the Four-Step Model and even more advanced Agent-Based Models (ABMs), have made significant contributions. Yet, they often struggle to represent the nuanced, adaptive, and sometimes irrational nature of human behavior. As cities become more complex and mobility patterns more dynamic, the limitations of these models become increasingly apparent.

Recent advancements in Large Language Models (LLMs) open a new frontier. These models, trained on vast amounts of human-generated data, demonstrate an unprecedented ability to mimic human reasoning, interpret context, and generate adaptive responses.

This raises a critical question for planners: Can LLMs help us model cities not just as systems but as lived experiences shaped by human decisions?

From Aggregate Models to Behavioral Realism

Sequential process of transportation planning. From “Inside the black box: Making transportation models work for livable communities,” by Beimborn, E., and Kennedy, R., 1996. In the public domain.

Historically, transportation modeling has evolved from aggregate approaches toward more disaggregate, behavior-sensitive frameworks.

The Four-Step Model, long considered the backbone of transportation planning, operates at an aggregate level. It estimates trip generation, distribution, mode choice, and route assignment based on zonal data. While robust and widely used, it assumes relatively stable and predictable behavior patterns.

In response to these limitations, ABMs were introduced. These models simulate individual agents (representing people or households) making decisions based on predefined rules. ABMs marked a significant shift toward behavioral realism, allowing planners to explore how individual actions collectively shape system outcomes.

However, even ABMs face critical constraints. Their behavioral rules are often rigid, requiring extensive calibration and large datasets. More importantly, they still struggle to fully capture bounded rationality which is the idea that human decisions are not perfectly rational but influenced by habits, perceptions, and incomplete information.

Enter LLM-Based Agents: A New Modeling Paradigm

This is where LLM-based agents introduce a paradigm shift.

Unlike traditional agents defined by fixed rules, LLM-based agents operate through natural language reasoning. They can interpret context, recall past interactions, and adapt decisions dynamically. In essence, they function as behaviorally rich proxies for human actors.

This approach enables a more flexible and expressive modeling framework:

• Agents can simulate daily activity patterns rather than isolated trips

• Decision-making can reflect context, memory, and preferences

• Behavioral responses can evolve over time through feedback loops

For urban planners, this represents a significant leap forward. Instead of encoding behavior mathematically, planners can define scenarios and constraints in natural language, allowing the model to “reason” through decisions.

Why This Matters for Urban Planning

From a city and regional planning perspective, the integration of LLM-based agents into transportation modeling is not just a technical improvement, it represents a shift in how planners understand and interact with urban systems. By moving beyond aggregate assumptions and rigid behavioral rules, this approach opens new possibilities for interpreting the complex relationship between people, space, and mobility.

src: https://www.edthefed.com/planning/documents/primer.htm

1. Stronger Link Between Behavior and Space

One of the long-standing challenges in urban planning is effectively linking individual behavior to spatial outcomes. Traditional models often treat travel as a function of aggregated variables such as population density or land use categories, which can obscure the underlying decision-making processes.

LLM-based agents, however, operate at the individual level, allowing planners to simulate how daily activities, preferences, and constraints shape mobility patterns. This makes it possible to observe how micro-level decisions, such as choosing when and why to travel, translate into macro-level spatial dynamics. As a result, planners can develop a more nuanced understanding of land use–transport interactions, enabling more informed decisions about zoning, accessibility, and urban form.

2. More Flexible Scenario Testing

Scenario analysis is central to planning practice, yet traditional modeling approaches often require significant time and technical effort to test even minor policy changes. Adjusting model parameters, recalibrating datasets, or rewriting code can be resource-intensive and slow down the decision-making process.

In contrast, LLM-based frameworks introduce a more flexible and iterative approach. By using natural language prompts, planners can quickly define and test alternative scenarios, such as new transit investments, pricing policies, or behavioral interventions, without restructuring the entire model. This significantly lowers the barrier to experimentation and enables planners to explore a wider range of possibilities in a shorter time frame.

Such flexibility is particularly valuable in rapidly changing urban environments, where planning decisions must respond to uncertainty and evolving mobility patterns.

3. Data Efficiency in Resource-Constrained Contexts

Access to high-quality, large-scale data is one of the main constraints in transportation planning, especially in developing or resource-limited contexts. Traditional models often depend on detailed surveys, origin-destination matrices, and extensive calibration processes, which can be costly and time-consuming.

LLM-based approaches offer an alternative by leveraging pre-trained knowledge and contextual reasoning capabilities. While they do not eliminate the need for data entirely, they have the potential to reduce dependency on highly granular datasets, enabling planners to work with more limited inputs.

This shift could democratize access to advanced modeling tools, allowing smaller municipalities or institutions with limited resources to engage in sophisticated scenario analysis and planning exercises.

4. Toward More Inclusive Planning

Urban systems are shaped by diverse populations with varying needs, preferences, and constraints. However, traditional models often rely on average behaviors or representative agents, which can mask important differences between social groups.

By simulating individual-level decision-making, LLM-based agents can better capture behavioral diversity across income groups, age cohorts, or lifestyle patterns. This opens up new opportunities to evaluate how different policies impact different segments of society.

From a planning perspective, this means moving toward more inclusive and equity-sensitive analyses, where the distributional effects of policies can be examined in greater detail. Ultimately, this contributes to more responsive and socially aware planning practices.

Rethinking Urban Mobility Through AI

The integration of LLM-based agents into transportation modeling signals more than a technical advancement. It represents a shift in how we conceptualize urban systems.

Cities are not just networks of roads and infrastructure; they are dynamic systems shaped by human behavior. By embedding more realistic representations of that behavior into our models, we move closer to planning cities that truly respond to human needs.

While challenges remain, the potential is clear:LLM-driven models could redefine how we simulate, evaluate, and ultimately design urban mobility systems.