Imagine a world where artificial intelligence helps design and optimize economic policies, leading to better social welfare and happiness. That’s exactly what a new preprint, LLM Economist, explores. By using large language models (LLMs) to simulate economic scenarios, the researchers have created a framework where a planner agent proposes tax policies, and a population of worker agents responds by choosing how much labor to supply based on their individual personas.
What’s fascinating is that all agents interact entirely through language, using in-context messages and JSON actions. The planner observes the behaviors and adjusts tax policy over time to maximize social welfare, without using gradient updates. Instead, it learns directly through repeated text-based interactions and the resulting societal and individual rewards.
The LLM Economist framework has some key contributions, including a two-tier in-context RL framework, persona-conditioned agent population grounded in real-world statistics, and emergent economic responses to policy changes. It’s a promising step towards using AI in economic modeling and mechanism design.
The researchers are seeking feedback on the viability of language-only RL architectures for economic modeling, the stability and interpretability of emergent agent behavior, and the broader implications for coordination and mechanism design with LLMs. You can read the full paper and explore the code on GitHub.
What do you think about using AI in economic policy design? Could this lead to more efficient and equitable societies?