TaxAgent: Large Language Model-Empowered Adaptive Taxation Optimizer

ACL ARR 2025 February Submission3951 Authors

Abstract: Economic inequality is a severe global issue. It intensifies disparities in education and healthcare, impairing social stability. Traditional systems such as the US federal income tax reduce inequality but they lack adaptability. Frameworks like the Saez optimal taxation introduce dynamic adjustments, but they rely on rigid economic assumptions and taxpayer homogeneity. This study introduces the TaxAgent, an integration of large language models (LLMs) with agent-based modeling (ABM) to design adaptive tax policy that accounts for taxpayer heterogeneity and moves beyond arbitrary assumptions. In our macroeconomic simulation, heterogeneous H-Agents (households) simulate real-world taxpayers and the TaxAgent (government) iteratively optimizes tax rates. Benchmarked against Saez optimal taxation, US federal income tax, and free-market system, TaxAgent achieves superior equality-productivity trade-offs and maintains a healthy economy with low unemployment and stable inflation. Two behavioral experiments further suggest that H-Agents better simulate real human decision-making compared to rule-based models. This research provides a novel taxation solution and a scalable framework for fiscal policy evaluation, demonstrating the potential of LLMs in addressing social challenges.

Paper Type: Long

Research Area: Computational Social Science and Cultural Analytics

Research Area Keywords: NLP tools for social analysis, human behavior analysis

Contribution Types: NLP engineering experiment

Languages Studied: English,

Submission Number: 3951