Research map / Multi-agent systems and games / Human-AI and goal-conditioned
EconGym: A Scalable AI Testbed with Diverse Economic Tasks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces EconGym, a modular testbed with 11 heterogeneous economic role types and 25+ tasks for training and benchmarking AI agents (RL, LLM, BC, rule-based) on economic policy problems, scaling to 10k agents.
Problem
Existing AI economic simulation platforms are narrow in scope (mostly single-domain taxation or market tasks), lack cross-domain multi-government coordination, and have limited agent heterogeneity and scale.
Approach
A modular platform that composes heterogeneous economic roles (OLG and Ramsey households, fiscal, central bank, pension authorities, commercial/non-profit banks, firm market structures) into Markov game environments where users assign diverse agent algorithms (RL, LLM, behavior cloning, economic rules, expert rules, real data) to each role.
Key results
- RL pension agent sustains the pension fund for about 165 years, nearly 100 years longer than the real-data agent (RA=67).
- Hybrid configuration Saez tax + DDPG + LLM achieves the highest GDP (3.06e+10) and lowest Gini (0.32) among 33 configurations, while pure LLM-only (ID-21) and PPO-only (ID-22) configurations underperform.
- Simulation with N=10k agents runs at 0.016 ms per agent per step for OLG; Wasserstein distance to real distributions decreases as population grows from 10 to 100,000.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add a second government policy role, using real-data policies for fiscal and central-bank authorities | Single-government coordination setup · Multi-government coordination | GDP: 2.23e+10 -> 2.33e+10 | Helped |
| Use Saez tax with DDPG for the pension authority and LLM for the central bank | Multi-government coordination · Multi-government coordination | GDP: 3.06e+10 | Helped |
| Use LLM-only government agents | Multi-government coordination · Multi-government coordination | GDP: 1.98e+09 | Hurt |
| Use PPO-only government agents | Multi-government coordination · Multi-government coordination | Welfare: -8.91e+04 | Hurt |
| Raise retirement age from 60 to 70 | Aging-pension scenario · Aging-pension | Pension fund longevity: 15 years | Helped |
| Raise retirement age | Aging-pension scenario · Aging-pension | Per capita utility | Hurt |
| Increase the simulated population size | EconGym population simulation · Realism evaluation | Wasserstein Distance | Helped |
Surprising findings
- Pure AI policies (LLM-only and PPO-only across all governments) perform far worse than hybrid AI plus economic-rule configurations.
- Ramsey model per-agent step time rises at N=100k, and multi-government combinations can produce conflicting outcomes (Fiscal + Central Bank + Pension is weaker than subsets).
Limitations
- Government interventions are not calibrated to real-world policy responses due to scarce datasets.
- Agent scale is limited to about 10k, far from the millions needed for population-level macro simulation.
- Core set of roles only; institutions like local governments, insurance, and international trade are not yet modeled.
Experimental setup
Datasets: 2022 Survey of Consumer Finances CDC 2022 mortality rates U.S. birth rates
Benchmarks: EconGym multi-government coordination tasks 2022 Survey of Consumer Finances realism comparison
Metrics: GDP consumption welfare Gini pension fund longevity Wasserstein distance step time
Baselines: AI Economist EconoJax R-MABM ABIDES-Economist TaxAI EconAgent AgentSociety
Builds on: AI Economist TaxAI EconAgent Saez tax Taylor rule
Closest papers by meaning
- EcoGym: Evaluating LLMs for Long-Horizon Plan-and-Execute in Interactive Economies (NeurIPS 2026 · similarity 0.87)
- Economy of Minds: Emerging Multi-Agent Intelligence with Economic Interactions (NeurIPS 2026 · similarity 0.85)
- TwinMarket: A Scalable Behavioral and Social Simulation for Financial Markets (NeurIPS 2025 · similarity 0.85)
- Position: Social Environment Design Should be Further Developed for AI-based Policy-Making (ICML 2024 · similarity 0.84)
- When AI Agents Compete for Jobs: Strategic Capabilities and Economic Dynamics of AI Labour Markets (ICML 2026 · similarity 0.83)
- InvestESG: A multi-agent reinforcement learning benchmark for studying climate investment as a social dilemma (ICLR 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.