acceptodds
Withdrawn from ICLR 2027

SimCity: Multi-Agent Urban Development Simulation with Rich Interactions

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Large Language Models (LLMs) open new possibilities for constructing realistic and interpretable macroeconomic simulations. We present SimCity, a multi-agent framework that leverages LLMs to model an interpretable macroeconomic system with heterogeneous agents and rich interactions. Unlike classical equilibrium models with limited heterogeneity, or traditional agent-based models (ABMs) that rely on hand-crafted decision rules, SimCity enables flexible, adaptive behavior with transparent natural-language reasoning. Within SimCity, four core agent types (households, firms, a central bank, and a government) make decisions and participate in a frictional labor market, a heterogeneous goods market, and a financial market. Furthermore, a Vision–Language Model (VLM) determines the geographic placement of new firms based on a virtual city map, allowing us to study both macroeconomic regularities and urban expansion dynamics within a unified environment. To evaluate the framework, we compile a checklist of canonical macroeconomic phenomena, including price elasticity of demand, Engel’s Law, Okun’s Law, the Phillips Curve, etc., and show that SimCity effectively exhibits these canonical patterns. Using our framework, we study novel economic shocks that are difficult to analyze in existing models and examine their macroeconomic implications qualitatively, yielding coherent results.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Macroeconomic, Social Simulation, Large Language Models, Multi-Agent System

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_StWhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces "SimCity," a multi-agent macroeconomic and urban simulation framework powered by large language models (LLMs). It models economic entities, such as households, firms, the government, and the central bank, as agents with heterogeneous reasoning capabilities. The framework incorporates markets for goods, labor, and finance, alongside mechanisms for taxation, transfer payments, and investment pools. It also features a visual urban map and a location-selection mechanism driven by vision-language models (VLMs) to simulate urban expansion and stable development in distinct phases.

Critical strengths

By modeling various economic entities as LLM-based agents, the framework expands the limited decision-making space found in previous LLM-driven macroeconomic simulations, providing households, firms, the government, and the central bank with a rich set of action interfaces.

It integrates structured market interactions with a visual urban map, utilizing vision-language models to determine business locations, thereby incorporating spatial layout and urban expansion directly into the simulation process.

Critical weaknesses

The simulation is qualitative rather than quantitative in nature. It lacks calibration against real-world macroeconomic data, making it unsuitable for precise forecasting or practical policy formulation.

The heavy reliance on LLMs introduces challenges such as high computational costs, hallucinations, inherent randomness, and the compounding of errors across multiple agents; these factors limit reproducibility, scalability, and the feasibility of large-scale, long-duration simulations.

The financial system is significantly simplified, lacking complex instruments like bonds, stocks, and derivatives, and investment decisions are not based on rigorous mathematical optimization, which may constrain the framework's ability to capture complex market dynamics.

Questions for authors

Please refer to the weaknesses.

Reviewer_55r4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

SimCity is an LLM-driven multi-agent macroeconomic simulation with households, firms, a government, and a central bank, + a VLM-based spatial module for city growth. The evaluation argues that the simulator reproduces several canonical macroeconomic regularities and can be used to study stylised exogenous shocks such as price shocks and automation shocks.

Critical strengths

-framework is ambitious and easy to understand. Compared with prior LLM-agent economy papers, it covers a broader set of agents and interactions -paper is readable and the implementation effort appears substantial. It is plausible that this could become a useful sandbox for qualitative exploration.

Critical weaknesses

-experimental validation is much weaker than the claims. Most results amount to showing that the simulator exhibits well-known stylized facts, but many of these patterns could be induced by the hand-designed mechanics already built into the environment -comparison standard is not strong enough for ICLR. The system is not quantitatively calibrated, uses only up to 1000 households, and the main shock analyses are qualitative case studies

Questions for authors

1.Which headline regularities still hold after ablating the specific hand-crafted mechanisms that most directly encode them, for example the Taylor rule, essential-goods constraints, or the production/investment rules? 2.Can you provide a stronger validation against real data or at least against externally calibrated targets, rather than mainly checking for stylized-sign consistency?

Reviewer_5dRWOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents SimCity, an LLM-driven multi-agent macroeconomic simulation. Households, firms, a government, and a central bank interact through three markets: a frictional labor market, a goods market with 44 heterogeneous goods derived from OECD input–output tables, and a simple financial market. A VLM-based investment pool decides where new firms are placed on a rendered city map, which couples macroeconomic dynamics with urban expansion. The authors evaluate the framework against a checklist of stylized facts and compare it with LEN, CATS, AI-Economist, EconAgent, and FRED data. They also run exploratory price-shock and automation-shock experiments.

Critical strengths
  1. The environment is substantially richer than prior LLM-based macro simulations such as EconAgent. Firms are LLM-driven and make pricing, hiring, and investment decisions, and the model adds an IO-table-based production network, taxation, and credit.
  2. The heterogeneous goods market makes it possible to test regularities that single-good frameworks cannot address, such as Engel’s law and good-specific elasticities.
  3. Coupling macroeconomic simulation with spatial urban development is an interesting and relatively unexplored direction.
  4. The paper is transparent. It provides full prompts, environment details and cost estimates, and reports robustness checks across seeds and an alternative backbone.
Critical weaknesses
  1. Limited novelty and heavy hand-crafted design. The framework mainly extends EconAgent with more agent types and markets, and the same four-agent setup already exists in ABIDES-Economist. Many outcomes rely on manual scaffolding rather than on the agents’ reasoning: The two-phase protocol does not reflect how real cities develop. Households immigrate up to a fixed cap, and the population is then closed with no further migration. Prompts contain explicit behavioral instructions. For example, Listing 4 tells households to shift spending from food to luxury goods when income rises, to adjust needs according to price changes, and to invest when ROI exceeds the interest rate.
  2. The evidence for the stylized facts is not convincing.
    • Okun’s law. Output is a direct Cobb–Douglas function of effective labor, so the correlation is largely mechanical. A random or rule-based work policy as a control would clarify this.
    • Engel’s law. The pattern is explicitly instructed in the household prompt, so it is not emergent. In Fig. 2 the distribution is also highly imbalanced: many near-zero-income households have an Engel coefficient of exactly 1, and the fit seems to be driven by a few high-income households.
    • Law of demand. Fig. 4 uses “Needs”, which the prompt tells agents to adjust with price changes, instead of realized purchases.
  3. The claimed advantage of LLM agents is not demonstrated. On the regularities that can be compared, SimCity is not stronger than the baselines in Okun’s law (r = −0.53, versus −0.81 for LEN, −0.78 for EconAgent and −0.72 for FRED), Beveridge curve (r = −0.52, versus −0.80 for CATS), and Phillips curve (on par with EconAgent, and it drops to −0.24 with gpt-5). Neither the main text nor the appendix analyzes how demographics, persona traits or the reasoning process affect the results. There is also no ablation that replaces LLM decisions with heuristic rules in the same environment, so the LLM’s contribution cannot be isolated.
  4. The central bank is not an LLM agent. The abstract, Contribution 1, the opening of Sec. 4 and the conclusion all claim four LLM-driven roles. However, Sec. 4.4 and App. C.5 implement the central bank as a Taylor rule with fixed smoothing, and Table 4 includes no central-bank prompt.
  5. The VLM module plays a limited role. The text-only ablation receives only a coordinate list. The linear layout may come from the prompt not specifying map boundaries rather than from the lack of vision.
  6. The experiment comparison is narrow and unfair. Baselines are compared only on the Phillips, Okun and Beveridge relations, all of which involve unemployment. Every other checklist item is N/A for the baselines, and the shock experiments have no baselines at all. Data frequencies also differ (quarterly vs. annual, footnote 2), so the reported r and p values are not directly comparable. The paper should compare with or discuss more recent and closely related work. I did a simple search, and here are related works: ABIDES-Economist (Dwarakanath et al., 2024), EconGym (Mi et al., 2025; cited but not compared), the hierarchical LLM macro framework at AAMAS 2025, and LLM Economist (Karten et al., 2025).
  7. The presentation is poor.
    • Fig. 6 appears before Fig. 5.
    • Results for the regularities are scattered across Figs. 2–4 and the appendix.
    • The text refers to “Figure 2 (top-right)”, but Fig. 2 is a single row.
Questions for authors
  1. Do the regularities still hold when the explicit behavioral instructions are removed from the prompts?
  2. Can you provide the results if simply replace LLM agents with heuristic functions?
  3. Sec. 5.1 states up to 1000 households, but the population in Fig. 6 plateaus around 200 at about month 16, even though phase 1 lasts 36 steps. Which configuration produced the main results?
  4. Why do the baseline correlations for FRED, CATS and LEN differ between Fig. 3 and Fig. 9, when only SimCity’s backbone changes?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.