LLM agents generate context-based decisions like humans, enabling individual simulation in economic experiments beyond rule-based agents.
Large language model (LLM)-driven agents exhibit human-like decision-making capabilities, enabling them to generate context-based decisions that go beyond the rigid frameworks of traditional rule-based systems . These agents can simulate complex economic behaviors such as employment, consumption, budget allocation, and responses to taxation and pricing, adapting dynamically to changing economic environments . Unlike rule-based agents, LLM agents demonstrate autonomous reasoning, perception, reflection, and action, allowing for more realistic and flexible simulations of individual behavior in socio-economic contexts .
Studies show that LLM agents can replicate key aspects of human economic behavior, including risk preferences, time preferences, and social decision-making in games like the ultimatum game and public goods game . They display bounded rationality rather than perfect rationality, mirroring human tendencies in market settings, such as trend-following in positive feedback markets and slower convergence to equilibrium prices . Furthermore, when endowed with personas derived from behavioral datasets, LLM agents can align their decisions more closely with real human patterns, enhancing their validity in behavioral economics research .
In large-scale simulations, LLM agents interact within synthetic societies that include urban, social, and economic spaces, where they adjust work hours, consumption budgets, and mobility based on dynamic factors like income, prices, and social influences . These agents can also form emergent social structures, exhibit individuality, and evolve behaviors through repeated interactions, making them powerful tools for modeling socio-economic systems .
The integration of psychological theories—such as Maslow’s Hierarchy of Needs and the Theory of Planned Behavior—into agent design further enhances their human-like qualities, enabling simulations of emotional states, motivations, and cognitive processes that influence economic choices . As a result, LLM-powered agent-based models offer a scalable, cost-effective alternative to traditional laboratory experiments for studying human behavior in economic and social systems .
This research introduces a framework for utilizing Large Language Models (LLMs) as autonomous agents capable of simulating human-like decision-making within socio-economic environments. Unlike traditional economic models that rely on rule-based agents or strict rational actor assumptions (homo economicus), this approach leverages the generative capabilities of LLMs to produce context-dependent behaviors. By instantiating agents with specific personas and backgrounds, the model allows for individual simulation where agents interpret scenarios, weigh social cues, and make decisions that closely mirror the nuances and heuristics of human reasoning rather than following pre-programmed utility functions.
The key contribution of this work is the validation of LLMs as viable proxies for human subjects in economic experiments, bridging the gap between behavioral economics and computational simulation. This matters because it enables researchers to conduct scalable, high-fidelity experiments that capture the complexity of human interaction—such as trust, cooperation, and irrationality—without the logistical constraints and costs of recruiting human participants. By moving beyond rigid rule-based systems, this socio-economic model offers a powerful new tool for policy testing, market dynamics prediction, and the study of emergent social phenomena in a controlled yet realistic digital environment.
This paper introduces a framework for modeling socio-economic behavior using Large Language Model (LLM) agents, which generate context-dependent decisions akin to human reasoning. Unlike traditional rule-based or reinforcement learning agents, these LLM agents simulate individual-level economic behavior in dynamic environments, enabling richer, more nuanced experiments. The work leverages the emergent capabilities of LLMs—such as understanding natural language, reasoning under uncertainty, and adapting to social norms—to capture human-like decision-making in economic contexts, including cooperation, competition, and strategic interaction.
The paper’s key contributions include: 1. A novel agent architecture that integrates LLMs with economic modeling, allowing for flexible, interpretable simulations of human behavior. 2. Empirical validation through experiments in classic economic games (e.g., ultimatum games, prisoner’s dilemma) and market mechanisms, demonstrating how LLM agents reproduce observed human patterns while incorporating novel behaviors. 3. Scalability and adaptability, as the model can be fine-tuned for domain-specific economic scenarios without full retraining, making it a practical tool for policy testing and behavioral economics research.
This work matters because it bridges the gap between computational modeling and behavioral economics, offering a powerful new tool for studying complex socio-economic systems. By enabling more realistic simulations of human decision-making, it could advance fields like autonomous systems, algorithmic policy design, and AI-driven economic forecasting, while also raising important questions about the ethical and theoretical implications of AI-mediated economic behavior.