EconGrowthAgent: Economic Growth Simulation based on LLM Agents and Growth Theory

ICLR 2026 Conference Submission15598 Authors

Keywords: economics, agent-based models, simulation, large language model

TL;DR: We propose an LLM-based economic simulation.

Abstract:

Economics has long sought to understand economic phenomena to enrich human society, yet it has been constrained by being "a discipline where experiments cannot be conducted'' due to ethical and practical limitations.
While Agent-Based Models (ABM) offer a computational alternative, modeling complex human decision-making remains fundamentally challenging.
Recent large Language Models (LLMs) provide new capabilities for this modeling.
We present $\textbf{EconGrowthAgent}$, the first LLM-based ABM that simulates economic growth---the most essential phenomenon in economics.
Our approach decomposes macroeconomic growth theory into a micro-level dynamic model, where decisions of LLM-based economic agents interact and evolve.
Through 25-year simulations with 100 agents, we demonstrate that EconGrowthAgent reproduces economic growth and related key macro phenomena while enabling seamless micro-to-macro analysis.
We further demonstrate its value by simulating a counterfactual scenario difficult to explore in reality -- an approaching civilization-ending asteroid.
As a "laboratory for economic experiments," EconGrowthAgent advances our understanding of economic phenomena and propels economics forward.

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 15598