💻 computer science

SimCity: Multi-Agent Urban Development Simulation with Rich Interactions

This paper introduces SimCity, a novel multi-agent simulation framework that leverages Large Language Models and Vision-Language Models to create an interpretable, adaptive macroeconomic environment with heterogeneous agents and rich interactions, successfully reproducing canonical economic laws and urban dynamics without relying on hand-crafted decision rules.

Original authors: Yeqi Feng, Yucheng Lu, Hongyu Su, Yixin Tao, Tianxing He

Published 2026-04-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yeqi Feng, Yucheng Lu, Hongyu Su, Yixin Tao, Tianxing He

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are the mayor of a brand-new, digital city. But instead of just moving pieces around on a board, you've hired a team of incredibly smart, talking robots (powered by advanced AI) to actually live there. These robots aren't just following a rigid script like "if X happens, do Y." Instead, they have personalities, dreams, fears, and the ability to reason through complex problems just like real humans do.

This is the core idea behind SimCity, a new research project that uses Large Language Models (LLMs) to simulate a whole economy.

Here is a breakdown of how it works, using some fun analogies:

1. The Cast of Characters (The Agents)

In traditional economic models, everyone is often treated as a "representative agent"—basically, a generic clone of an average person. In SimCity, everyone is unique. Think of it like a massive, living role-playing game (RPG) where every NPC (Non-Player Character) has their own backstory.

The Households: These are the families. Some are young and eager to spend; others are older and saving for retirement. Some are great at coding, others are great at cooking. They need to buy food, pay rent, and find jobs.
The Firms: These are the businesses. They aren't just static factories; they are dynamic entities that decide what to make, how much to charge, and who to hire based on what's happening in the city.
The Government: The mayor's office. They collect taxes, build public parks, and hand out welfare checks to keep things running smoothly.
The Central Bank: The money manager. They adjust interest rates (the cost of borrowing money) to keep inflation (rising prices) from getting out of hand.

2. The Brain: Why LLMs?

In old-school simulations, if a family needed money, the computer would run a simple math formula: Income - Expenses = Savings. If the result was negative, the family went bankrupt. It was rigid and predictable.

In SimCity, the "brain" of every character is an LLM (like the technology behind ChatGPT).

The Analogy: Imagine a family meeting. Instead of a calculator, the family sits around a table and talks. "Well, the price of bread went up, so maybe we should eat out less this month," says the dad. "But the kids need new shoes, so maybe we should skip the vacation," says the mom.
The AI simulates this conversation. It reads the news, checks the bank account, and then decides what to do based on human-like reasoning. This allows for heterogeneity (everyone is different) and adaptability (they can learn and change their minds).

3. The World: A Visual City

The simulation isn't just a spreadsheet; it's a visual map.

The Vision: A special type of AI (a Vision-Language Model) looks at a map of the city. It decides where to build new factories and houses.
The Result: Just like in real life, the AI naturally organizes the city. It puts residential areas near the center and factories on the outskirts, mimicking how real cities grow. If you remove the visual map and only give the AI text descriptions, the city becomes a weird, straight line of buildings—proving that "seeing" the city helps the AI understand space better.

4. The Test: Does It Make Sense?

The researchers didn't just want to see if the robots could talk; they wanted to see if the economy they created followed the laws of economics. They tested the simulation against famous economic "rules of thumb" (like the Phillips Curve, which links unemployment and inflation, or Engel's Law, which says rich people spend a smaller percentage of their income on food).

The Verdict: The simulation passed the test! The AI-driven city naturally developed these complex economic patterns without anyone explicitly programming them to do so. The robots figured out that when unemployment goes up, inflation often goes down, just like in the real world.

5. The "What If" Scenarios

Because the agents are so flexible, the researchers can ask "What if?" questions that are hard to answer with traditional math models.

The Automation Shock: What if all truck drivers suddenly became obsolete because of self-driving cars? The simulation showed a massive gap opening up between workers with "IT skills" and those with "physical labor skills," creating a new kind of wealth inequality.
The Price Shock: What if the price of electricity suddenly doubled for a month? The simulation showed that prices eventually settle back down, but it takes time (a concept called "price stickiness").

Why Does This Matter?

Think of SimCity as a flight simulator for the economy.

Old Way: Economists build a model based on simplified math. It's like flying a plane with a map that only shows straight lines.
New Way (SimCity): Economists can now run a simulation with thousands of unique, thinking agents. It's like flying a plane in a full flight simulator with realistic weather, traffic, and passenger behavior.

This allows policymakers and researchers to test new ideas (like a new tax or a sudden tech boom) in a safe, virtual environment before trying them in the real world. It's a bridge between the rigid math of economics and the messy, unpredictable reality of human behavior.

1. Problem Statement

Traditional macroeconomic modeling faces a dichotomy between Dynamic Stochastic General Equilibrium (DSGE) models and Agent-Based Models (ABMs):

DSGE models rely on representative agents and explicit optimization, offering mathematical elegance but failing to capture agent heterogeneity and complex behavioral dynamics.
Traditional ABMs accommodate heterogeneity but typically depend on hand-crafted, rigid decision rules that lack adaptability and interpretability.
Existing LLM-driven simulations (e.g., EconAgent) have been limited in scope, often restricting agent-environment interactions to simplified variables (e.g., consumption share only) and lacking spatial context or complex market structures (e.g., heterogeneous goods, financial markets).

The paper addresses the need for a macroeconomic simulation framework that combines heterogeneous agents, rich interactions, spatial dynamics, and transparent, interpretable reasoning to study both established economic laws and novel shock scenarios.

2. Methodology: The SimCity Framework

SimCity is a multi-agent system powered by Large Language Models (LLMs) and Vision-Language Models (VLMs). It simulates an urban economy through three core layers: Environment, Interaction Protocol, and Agents.

A. Environment and Simulation Phases

The simulation runs in monthly time steps and is divided into two phases:

Move-in Phase (36 steps/3 years): New households (immigrants/newborns) enter the city. This phase explicitly models urban expansion and the evolution of spatial economic activity.
Development Phase (144 steps/12 years): The population is fixed. The system operates in a stable state to study macroeconomic regularities without transient noise.

The environment features three markets:

Goods Market: Firms sell qualitatively differentiated goods (44 categories, e.g., food, energy, healthcare) to households and other firms.
Labor Market: Firms post jobs; households apply based on skill matching.
Financial Market: A central bank manages interest rates; an investment pool facilitates equity financing for new firm formation.

A Visualized Map renders the city, showing residential, productive, and government buildings. A Vision-Language Model (VLM) determines the geographic placement of new firms to ensure realistic urban clustering.

B. Agent Types

Four distinct agent types are implemented as LLM-driven entities:

Households: Heterogeneous agents with unique profiles (age, education, skills, preferences). They make decisions on consumption bundles (adhering to essential needs), labor participation (job acceptance/resignation), housing (renting), and financial activities (saving, borrowing, investing).
Firms: Agents instantiated from templates representing 44 industrial sectors. They utilize a Cobb-Douglas production function ( $Y = A L^{1-\alpha} K^\alpha$ ) where labor efficiency depends on skill matching. Firms decide on output quantity/price, hiring/firing, wage setting, and capital investment (borrowing or equity financing).
Government: A single institutional agent that collects progressive income tax and Value-Added Tax (VAT). It manages fiscal policy through public spending (building infrastructure), Universal Basic Income (UBI) transfers, and debt management.
Central Bank: A single institutional agent that sets policy interest rates using a modified Taylor Rule, responding to inflation gaps and output gaps, with an interest rate smoothing mechanism to mimic real-world policy persistence.

C. Interaction Protocol

Reasoning: Agents receive structured prompts containing personal reports, city-wide statistics (prices, unemployment, interest rates), and environmental observations.
Action Execution: Agents use Function Calling to interact with the environment. The LLM outputs actions (e.g., find_job, set_price, adjust_tax_bracket) in JSON format, which the framework verifies and executes.
Memory: Agents maintain structured memory of their history and traits to ensure consistent decision-making over time.

3. Key Contributions

LLM-Driven Macroeconomic Simulation with Rich Interactions: The framework integrates LLMs to model four distinct economic roles with high heterogeneity. Unlike previous works, it supports complex interactions across labor, goods, and financial markets, including taxation and equity financing.
Visualized Urban-Style Simulation Environment: SimCity introduces a spatial dimension where a VLM guides the geographic placement of firms. This allows the study of urban expansion dynamics and the emergence of spatial structures (e.g., mono-centric city models) that text-only models cannot capture.
Systematic Evaluation and Novel Shock Analysis: The authors compiled a checklist of canonical macroeconomic "stylized facts" to validate the model. Furthermore, the framework enables the simulation of novel shocks (e.g., automation replacing physical labor, exogenous price shocks) that are difficult to model in traditional DSGE or rule-based ABMs.

4. Experimental Results

The authors evaluated SimCity against baseline models (traditional ABMs like LEN and CATS, deep RL models like AI-Economist, and real-world data from FRED).

A. Emergence of Macroeconomic Regularities

SimCity successfully reproduced several established economic laws:

Phillips Curve: Demonstrated the inverse relationship between unemployment and inflation.
Okun's Law: Showed the negative correlation between changes in unemployment and GDP growth.
Beveridge Curve: Captured the negative relationship between job vacancy rates and unemployment.
Engel's Law: Validated that the proportion of income spent on food decreases as income rises (requiring a heterogeneous goods market).
Law of Demand: Estimated Price Elasticity of Demand (PED) across 44 goods, correctly identifying necessities as inelastic and luxury goods as elastic.
Business Cycle Volatility: Replicated the real-world phenomenon where investment volatility exceeds consumption volatility.

B. Spatial Dynamics

During the "move-in" phase, the VLM-driven agents naturally formed a mono-centric city structure (residential areas near the center, production on the periphery), consistent with the Alonso-Muth-Mills (AMM) model. Ablation studies confirmed that removing the visual module resulted in unrealistic, linear city layouts, proving the necessity of visual grounding for spatial logic.

C. Response to Shocks

Price Shocks: When exogenous price impulses were applied, the system exhibited price stickiness, with prices gradually returning to equilibrium over time.
Automation Shock: Simulating a shift where physical labor is replaced by IT skills resulted in a widening income gap between skilled and unskilled workers, highlighting distributional consequences of technological change.

D. Robustness

Experiments with different random seeds and alternative LLM backbones (including GPT-5) confirmed that the observed macroeconomic regularities are robust and reproducible.

5. Significance and Limitations

Significance:
SimCity represents a paradigm shift in economic simulation by moving from rigid, rule-based agents to adaptive, reasoning-based agents. It bridges the gap between the interpretability of LLMs and the structural rigor of macroeconomic theory. The framework provides a testbed for studying complex, non-linear economic phenomena and urban planning dynamics that are currently intractable in standard models.

Limitations:

Abnormal Behavior: LLMs may occasionally generate aberrant decisions (e.g., extreme price changes), requiring heuristic checks.
Financial Complexity: The current model lacks complex financial instruments (bonds, derivatives) and does not use mathematical optimization for investment strategies.
Calibration: The model is not quantitatively calibrated to specific real-world datasets for prediction purposes; it is designed for qualitative exploration and mechanism analysis.

In conclusion, SimCity offers a powerful, interpretable platform for exploring macroeconomic dynamics, urban development, and the impact of policy shocks, leveraging the reasoning capabilities of modern AI to simulate human-like economic behavior.