Pith. sign in

REVIEW 6 major objections 5 minor 1 cited by

EconGym: A Scalable AI Testbed with Diverse Economic Tasks

T0 review · 6 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read EconGym claims that a modular Markov-game testbed built from 11 reusable economic roles supports more than 25 economic tasks, and that hybrid policies blending classical economic rules with AI outperform pure learning agents in complex…

desk verdict Useful platform, but the central empirical claim does not hold up: the main benchmark table is internally inconsistent and the realism check is quasi-circular. read the letter →

arxiv 2506.12110 v1 pith:7LUDB436 submitted 2025-06-13 econ.GN cs.AIq-fin.EC

Qirui Mi , Qipeng Yang , Zijun Fan , Wentian Fan , Heyang Ma , Chengdong Ma , Siyu Xia , Bo An
show 2 more authors
Jun Wang Haifeng Zhang
This is my paper
classification econ.GNcs.AIq-fin.EC
keywords economicsimulationmulti-agentreinforcementlearninglargelanguagemodelspolicycoordinationpensionmonetaryMarkovgamesagent-basedbenchmarking
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

EconGym is a testbed for training and evaluating AI agents on economic policy problems. The paper argues that a small set of well-specified economic roles—households, governments, banks, and firms, each with multiple variants—can be composed like building blocks into a shared multi-agent decision environment instead of building one bespoke simulator per question. This composition demonstrably covers 25 example tasks, from retirement-age and pension policy to fiscal, monetary, and mixed policy coordination. The paper reports that adding roles and algorithm families enlarges the policy space that agents can explore, and that configurations which embed classical economic prescriptions alongside learning agents achieve better GDP, welfare, and equality outcomes than pure learning baselines. A reader should care because the platform offers a single scalable arena where different AI and economic decision strategies can be compared under the same dynamic economic conditions.

What carries the argument

The load-bearing mechanism is role-type composition inside a Markov game, a multi-agent decision framework in which each agent observes its own state, chooses an action, and receives a reward, with the environment transition governed by economic constraints. EconGym models each economic actor as a modular agent—infinitely lived households, overlapping-generation households, fiscal and monetary and pension authorities, passive or commercial banks, and firms under four market structures—each with a private observation space, action space, and reward function, and defines transitions from the standard economic constraints of those roles, including budget constraints, market clearing, a no-arbitrage condition, and policy rules. This machinery lets users construct a task by selecting roles and algorithms rather than writing a new environment, and it simultaneously generates dynamic trajectories and macroeconomic indicators (GDP, consumption, welfare, inequality, pension fund balance) that support both economic analysis and AI policy optimization.

What would settle it

Perturb the parameters of the embedded policy rules—for example vary the inflation-response coefficient of the interest-rate rule and the elasticity of the optimal-income-tax formula by plus or minus 20%—and rerun the fiscal, monetary, and pension coordination scenarios; if the ranking of hybrid over pure-AI configurations on GDP, welfare, and Gini is not preserved, or if markets fail to clear under small perturbations, the composition claim is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that economic modeling can be decomposed into a small number of reusable role types with explicit interactions, and that recombining those roles yields a flexible, extensible testbed for AI-driven economic analysis. It implements 11 heterogeneous role types across individuals (infinitely lived and overlapping-generation households), governments (fiscal, monetary, pension authorities), banks (passive non-profit platforms and profit-seeking commercial banks), and firms (perfect competition, monopoly, oligopoly, monopolistic competition), and converts each role into an agent with defined observations, actions, and rewards. The experiments show that adding government agents and mixing algorithm families expands the set of policies that can be explored, that purely LLM-based or purely RL-based policies tend to fail in complex coordination tasks, and that hybrid configurations combining classical economic rules with RL or LLM agents consistently yield the highest GDP, highest welfare, and lowest inequality among the 33 configurations tested. The paper also reports that as the simulated number of households grows, the simulated distributions of consumption, labor, age, and wealth converge toward empirical distributions, while per-agent computation remains efficient at populations of 10,000.

Load-bearing premise

The load-bearing premise is that combining role models designed independently for different economic questions yields one coherent shared simulation, so that outcomes like GDP, welfare, and inequality carry the same meaning in every configuration being compared.

Editorial extensions

If this is right

  • New policy questions can be studied by recombining existing role types rather than building new environments, so the platform's demonstrated task count of 25 is a lower bound on what the design can support.
  • Pure LLM and pure RL policies, without structural economic priors, should be expected to underperform in high-dimensional multi-government settings, making hybrid configurations the likely benchmark winners for such tasks.
  • Populations of 100 to 10,000 agents are the practical operating range, where realism improves with population size before per-agent computation cost rises.
  • Coordinated fiscal, monetary, and pension policies can be either synergistic or conflicting depending on how their rule-based components interact, so multi-government designs require empirical benchmarking rather than assumed complementarity.
  • Replacing one rigid rule-based component of a policy stack with an adaptive learning agent can improve overall outcomes, suggesting a division of labor between economic rules and learning-based components.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • An implication left implicit is that the classical economic rules in hybrid configurations may be acting as prior knowledge or reward shaping for the learning agent; removing the rule while keeping the learning agent would isolate how much of the gain comes from each part.
  • The compositionality claim predicts that swapping one role type for another (for example, a monopolistic firm for a perfectly competitive firm) should shift prices, output, and welfare in economically predictable directions; a systematic grid of such swaps would test the building-block assumption more sharply than the six configurations reported.
  • The realism result rests on convergence toward one empirical dataset, so benchmarking the same scaling analysis against data from other countries or periods would separate genuine population effects from calibration to a single target.
  • The paper's own limitations section concedes that government interventions are not yet calibrated to real-world responses, so benchmark rankings should be read as comparisons internal to the simulator rather than quantitative forecasts.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

6 major / 5 minor

Summary. The manuscript introduces EconGym, a modular agent-based economic simulation testbed. It claims to support 11 heterogeneous economic role types, more than 25 composable tasks, multiple agent algorithms (RL, LLM, behavior cloning, economic rules, real-data replay), and scalability to 10k agents. The paper validates the platform through three experiments: pension policy optimization under aging, multi-government policy coordination, and realism/efficiency scaling. The headline empirical claim is that hybrid configurations combining AI with classical economic rules consistently outperform pure learning-based approaches in complex settings (Section 4.3 and Appendix A.1). The code is available online and the appendices contain full model specifications.

Significance. If the empirical claims were well supported, EconGym would be a useful contribution to the AI-for-economics testbed landscape: it integrates a wider set of economic roles than existing platforms, supports cross-domain policy coordination, and explicitly ships code and model details. The paper's compositional design, with clearly separated economic modeling and agent modeling layers, is a genuine strength. However, the current experimental evidence contains internal inconsistencies that prevent acceptance of the central empirical claims, so the significance is conditional on a substantial revision of the benchmark reporting and of the realism evaluation.

major comments (6)
  1. [Appendix A.1, Table 6, IDs 22–23] IDs 22 and 23 both list the same configuration, PPO/PPO/PPO, yet report terminal years 5 versus 54, GDP 1.50e+9 versus 1.18e+10, and welfare -8.91e+4 versus 9.96e+5. No seed count, standard deviation, or error bars are reported for any row of Table 6. This means the differences used in Section 4.3 and A.1 to rank configurations (e.g., ID-22 as a failed pure-RL baseline) cannot be distinguished from run-to-run variance or from a table error, and the central hybrid-superiority claim is not yet supported.
  2. [Appendix A.1, Table 6, IDs 24–25 and 26–29] IDs 24 and 25 differ only in the Central Bank algorithm (DDPG versus Taylor rule) yet report bit-identical consumption, average labor, GDP, welfare, and Gini; IDs 26 and 29 differ only in the Central Bank algorithm (Real data versus PPO) and are again bit-identical. These coincidences imply either that the stated algorithms are not actually varied in those simulation runs or that rows were copied. Because the conclusion that hybrid approaches (such as IDs 27 and 28) are best is derived from comparisons against rows like these, the evidence is internally inconsistent and must be resolved before any ranking can be trusted.
  3. [Section 4.4, Figure 6] The realism evaluation is quasi-circular. Section 4.1 states that households follow behavior-cloning policies trained on the 2022 Survey of Consumer Finances, and Figure 6 compares simulated population distributions against that same 2022 SCF as 'real data'. The observed decrease in Wasserstein distance as N grows is the mechanical effect of reduced sampling noise from a larger Monte Carlo sample; it does not independently validate the economic model or the cloned policy. The claim that 'larger populations enhance realism' is therefore not supported by this protocol. An out-of-sample comparison (e.g., a different survey wave, or holdout moments not used in calibration) would be needed.
  4. [Section 4.3, Figure 5 and Table 5] The conclusion that replacing the pension rule with an RL agent ('Fiscal + Central Bank + Pension (RL)') achieves the best overall results rests on a single trajectory. No error bars, multiple seeds, or statistical tests are reported anywhere in the paper, and given the internal inconsistencies in Table 6, this conclusion is not robust. At minimum, the benchmark needs multiple independent runs with mean and dispersion reported, and a statement of the termination criterion for each configuration.
  5. [Section 3.1 and Appendix C] EconGym composes independently designed economic models (Saez tax, Taylor rule, IMF pension rule, OLG/Ramsey households) into a single Markov game, but the paper provides no market-clearing verification, no sensitivity analysis for key parameters (e.g., the central bank weight λπ in Eq. (4), the tax curvature parameters ξ and ξa in Eq. (3)), and no check that the combined dynamics are not an artifact of arbitrary interaction rules. Without such checks, the simulated GDP, welfare, and Gini values in Tables 5–6 cannot be interpreted as meaningful economic outcomes. A concrete remedy would be to add a robustness section that perturbs key parameters and verifies that market-clearing conditions such as Eq. (21) and the no-arbitrage condition Eq. (12) hold in all reported runs.
  6. [Tables 5–6, 'Year' column] The 'Year' column is ambiguous: for some rows (e.g., ID-18, ID-22) the simulation terminates after 2–5 periods, while for others it runs for 70–80 periods. It is unclear what triggers termination (pension fund depletion, GDP collapse, numerical failure) and whether the reported GDP and welfare are terminal values, averages over the simulation horizon, or values at a fixed horizon. Comparing ID-22 (year 5) with ID-27 (year 79) on GDP is not meaningful if the time horizons differ for endogenous reasons, and the table should define this column explicitly.
minor comments (5)
  1. [Section 4.4 and Figure 7] The abstract and Section 4.4 emphasize scalability to 10k agents, but Figure 7 and Table 7 report results up to N=100,000; please reconcile the claimed scale with the presented benchmarks.
  2. [Section 4.2, Figure 4] Figure 4 labels the RL agent as 'DDPG' while the text refers generically to 'RL agents'; the training hyperparameters (learning rate, discount factor, network architecture, number of seeds) should be reported in the appendix for reproducibility.
  3. [Table 4] The rows for Perfect Competition firms and the Non-Profit Platform bank leave observation/action/reward cells blank; a short entry such as 'passive/price-taker' would make the table self-contained.
  4. [Appendix C.2, Eq. (4) and Table 12] The central bank objective includes the trade-off weight λπ and targets π* and g*, but no numerical values or calibration procedure are given; please specify these for reproducibility.
  5. [References] Reference [36] is formatted as 'Rui and Shi' while other references use full author names; please standardize the bibliography style.

Circularity Check

1 steps flagged · score 4.0 of 10

Realism claim is partly in-sample: BC household policies are trained on the 2022 SCF and then validated against that same SCF, so the decreasing Wasserstein distance with population size is partly a Monte Carlo artifact.

  1. fitted input called prediction [Sec. 4.1 (Behavior Cloning) and Sec. 4.4 (Realism vs. Efficiency, Fig. 6)]
    "In our experiments, individual households follow BC policies learned from the 2022 Survey of Consumer Finances data, enabling realistic micro-level behavior. ... We assess realism by comparing simulated population distributions to the 2022 Survey of Consumer Finances. ... As shown in Figure 6, WD consistently declines across consumption, labor, age, and wealth as N grows, validating that larger populations enhance realism in EconGym."

    The realism target is the same dataset used to fit the household behavior-cloning policies. A BC policy trained on the 2022 SCF will, by construction, generate samples whose empirical distributions converge to the fitted distribution as the number of simulated households grows; the reported decline in Wasserstein distance with N is therefore largely a law-of-large-numbers property of an in-sample fit, not an independent validation of EconGym's realism. The headline 'larger populations enhance realism' is thus partly forced by the choice of training and evaluation data.

full rationale

The platform construction and benchmark experiments are otherwise self-contained: the Markov-game formulations, economic-role taxonomies, and algorithm comparisons do not reduce to their inputs, and the self-citations (TaxAI, MF-LLM, DSMFG) are background context rather than load-bearing premises. The central hybrid-superiority claim rests on Table 6, which contains apparent data-quality problems (identical configurations with divergent outcomes, e.g., IDs 22/23, and bit-identical rows across different algorithms, e.g., IDs 24/25 and 26/29); those are correctness concerns, not circularity, so they do not raise the circularity score. The one genuine circular element is the realism evaluation: household policies are fit to the 2022 SCF and then compared to the same 2022 SCF to claim scale-dependent fidelity. This gives a partial circularity score of 4.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The platform itself is a software artifact, not a new theoretical entity. The free parameters are experimental choices (BC weights, RL hyperparameters, reward weights) that are not reported, limiting reproducibility. The axioms are standard economic modeling assumptions plus the implicit assumption that composing these models yields coherent dynamics.

free parameters (3)
  • Behavior cloning policy network weights = Not reported
    Household agents in experiments use BC policies trained on the 2022 Survey of Consumer Finances; the resulting network weights are fitted parameters that shape individual behavior. The paper does not report the architecture, training procedure, or final weights.
  • RL hyperparameters (e.g., learning rate, discount factor for PPO/DDPG) = Not reported
    The RL agents used in benchmark experiments (DDPG, PPO) require hyperparameters that are not specified in the paper, making exact reproduction impossible.
  • Central bank reward trade-off weight lambda_pi = Not stated
    The central bank reward combines inflation and growth deviations with a weight lambda_pi. The value is chosen by the authors but not reported in the main text or appendix tables.
assumptions (3)
  • domain assumption Standard economic models (utility maximization, CRRA preferences, Cobb-Douglas production, CES demand) are valid representations of household and firm behavior.
    Used throughout Section 3 and Appendix C to define agent utility, production, and market interactions. These are standard but are assumptions that may not hold in all economic settings.
  • domain assumption The 2022 Survey of Consumer Finances data is representative of U.S. household finances and is suitable both for training behavior cloning policies and for validating simulation realism.
    The realism evaluation (Section 4.4) compares simulated distributions to the same SCF data used to calibrate household policies. This assumes the SCF is a valid ground truth for both purposes.
  • domain assumption Composing independently designed economic roles into a single Markov game preserves economic consistency (e.g., no double counting, market clearing).
    Section 3.1 treats roles as 'universal building blocks' with 'explicit interactions', but no formal consistency check is provided for combined environments beyond individual model definitions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EconGym: A Scalable AI Testbed with Diverse Economic Tasks." pith.science (2026). https://pith.science/paper/7LUDB436

@misc{pith2026250612110,
  author       = {Pith},
  title        = {Pith review of: EconGym: A Scalable AI Testbed with Diverse Economic Tasks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7LUDB436}},
  note         = {Machine review of arXiv:2506.12110}
}
read the original abstract

Artificial intelligence (AI) has become a powerful tool for economic research, enabling large-scale simulation and policy optimization. However, applying AI effectively requires simulation platforms for scalable training and evaluation-yet existing environments remain limited to simplified, narrowly scoped tasks, falling short of capturing complex economic challenges such as demographic shifts, multi-government coordination, and large-scale agent interactions. To address this gap, we introduce EconGym, a scalable and modular testbed that connects diverse economic tasks with AI algorithms. Grounded in rigorous economic modeling, EconGym implements 11 heterogeneous role types (e.g., households, firms, banks, governments), their interaction mechanisms, and agent models with well-defined observations, actions, and rewards. Users can flexibly compose economic roles with diverse agent algorithms to simulate rich multi-agent trajectories across 25+ economic tasks for AI-driven policy learning and analysis. Experiments show that EconGym supports diverse and cross-domain tasks-such as coordinating fiscal, pension, and monetary policies-and enables benchmarking across AI, economic methods, and hybrids. Results indicate that richer task composition and algorithm diversity expand the policy space, while AI agents guided by classical economic methods perform best in complex settings. EconGym also scales to 10k agents with high realism and efficiency.

Figures

Figures reproduced from arXiv: 2506.12110 by the authors.

Figure 1
Figure 1. Overview of EconGym. Users define tasks by selecting economic roles and agent algorithms, generating dynamic multi-agent trajectories. These trajectories support economic analysis for the economics community and policy optimization for AI community. Built on rigorous economic theory and modular agent modeling, EconGym enables diverse and cross-domain economic tasks. Despite this progress, realizing the full potentia… view at source ↗
Figure 2
Figure 2. Workflow of EconGym. 3 The Structure of EconGym This section first introduces the workflow of EconGym, followed by its core components: economic modeling (§3.1) and agent modeling (§3.2). Workflow of EconGym. To solve a target economic problem, users follow two steps ( [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Evolution of economic indicators under varying retirement ages (60, 63, 65, 67, 70) in the [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Performance of agent algorithms (RL, LLM, rule-based, real-data) in pension policy [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Simulation outcomes under 6 heterogeneous combinations of fiscal, central bank, and [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Realism evaluation across scales (N=10 to 100,000): Top row compares consumption by [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 7
Figure 7. Figure 7: Step time comparison for OLG and Ramsey models across scales (N=10∼100k): (a) Total step time increases with N, with Ramsey scaling worse; (b) Per-agent step time decreases for both until N=10k, then rises for Ramsey. Efficient Simulation Up to 1K Agents. As the number…

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. CoRenew: A large language model agent-based policy simulation platform for multifamily residential redevelopment

    cs.MA 2026-07 conditional novelty 6.0 of 10

    CoRenew shows that LLM-based agents can reproduce the qualitative structure of real negotiation behavior and simulate the equity effects of redevelopment policies.

Reference graph

Works this paper leans on

50 extracted references · 39 canonical work pages · cited by 1 Pith paper

  1. [1]

    Provisional mortality data—united states, 2022.Morbidity and Mortality Weekly Report, 72(18):488, 2023

    Farida B Ahmad, Jodi A Cisewski, Jiaquan Xu, and Robert N Anderson. Provisional mortality data—united states, 2022.Morbidity and Mortality Weekly Report, 72(18):488, 2023

    work page 2022

  2. [2]

    Uninsured idiosyncratic risk and aggregate saving.The Quarterly Journal of Economics, 109(3):659–684, 1994

    S Rao Aiyagari. Uninsured idiosyncratic risk and aggregate saving.The Quarterly Journal of Economics, 109(3):659–684, 1994

    work page 1994

  3. [3]

    AI and Macroeconomic Modeling: Deep Reinforcement Learning in an RBC Model, March 2023

    Tohid Atashbar and Rui Aruhan Shi. AI and Macroeconomic Modeling: Deep Reinforcement Learning in an RBC Model, March 2023

    work page 2023

  4. [4]

    The impact of machine learning on economics

    Susan Athey. The impact of machine learning on economics. InThe Economics of Artificial Intelligence: An agenda, pages 507–547. University of Chicago Press, 2018

    work page 2018

  5. [5]

    MIT press, 1989

    Olivier Blanchard and Stanley Fischer.Lectures on Macroeconomics. MIT press, 1989

    work page 1989

  6. [6]

    Simu- lating the economic impact of rationality through reinforcement learning and agent-based modelling

    Simone Brusatin, Tommaso Padoan, Andrea Coletta, Domenico Delli Gatti, and Aldo Glielmo. Simu- lating the economic impact of rationality through reinforcement learning and agent-based modelling. In Proceedings of the 5th ACM International Conference on AI in Finance, pages 159–167, 2024

    work page 2024

  7. [7]

    Survey on large language model-enhanced reinforcement learning: Concept, taxonomy, and methods.IEEE Transactions on Neural Networks and Learning Systems, 2024

    Yuji Cao, Huan Zhao, Yuheng Cheng, Ting Shu, Yue Chen, Guolong Liu, Gaoqi Liang, Junhua Zhao, Jinyue Yan, and Yun Li. Survey on large language model-enhanced reinforcement learning: Concept, taxonomy, and methods.IEEE Transactions on Neural Networks and Learning Systems, 2024

    2024

  8. [8]

    Deep Reinforcement Learning in a Monetary Model, January 2023

    Mingli Chen, Andreas Joseph, Michael Kumhof, Xinlei Pan, and Xuan Zhou. Deep Reinforcement Learning in a Monetary Model, January 2023

    work page 2023

Show all 50 references
  1. [9]

    AI-driven Prices for Externalities and Sustainability in Production Markets, January 2023

    Panayiotis Danassis, Aris Filos-Ratsikas, Haipeng Chen, Milind Tambe, and Boi Faltings. AI-driven Prices for Externalities and Sustainability in Production Markets, January 2023

    2023

  2. [10]

    Saving and liquidity constraints, 1989

    Angus Deaton. Saving and liquidity constraints, 1989

    1989

  3. [11]

    Consumption and savings with large language model agents

    Michael R Douglas and Sergiy Verstyuk. Consumption and savings with large language model agents. Available at SSRN 4909749, 2024

    2024

  4. [12]

    Abides-economist: Agent-based simulation of economic systems with learning agents.arXiv preprint arXiv:2402.09563, 2024

    Kshama Dwarakanath, Svitlana Vyetrenko, Peyman Tavallali, and Tucker Balch. Abides-economist: Agent-based simulation of economic systems with learning agents.arXiv preprint arXiv:2402.09563, 2024

    2024 arXiv

  5. [13]

    Fiscal monitor: A fair shot, April 2021

    International Monetary Fund. Fiscal monitor: A fair shot, April 2021

    2021

  6. [14]

    Consumption over the life cycle.Econometrica, 70(1):47–89, 2002

    Pierre-Olivier Gourinchas and Jonathan A Parker. Consumption over the life cycle.Econometrica, 70(1):47–89, 2002

    2002

  7. [15]

    Deepham: A global solution method for heterogeneous agent models with aggregate shocks.arXiv preprint arXiv:2112.14377, 2021

    Jiequn Han, Yucheng Yang, et al. Deepham: A global solution method for heterogeneous agent models with aggregate shocks.arXiv preprint arXiv:2112.14377, 2021

    2021 arXiv

  8. [16]

    A multi-llm-agent-based framework for economic and public policy analysis

    Yuzhi Hao and Danyang Xie. A multi-llm-agent-based framework for economic and public policy analysis. arXiv preprint arXiv:2502.16879, 2025

    2025

  9. [17]

    Solving heterogeneous general equilibrium economic models with deep reinforcement learning, 2021

    Edward Hill, Marco Bardoscia, and Arthur Turrell. Solving heterogeneous general equilibrium economic models with deep reinforcement learning, 2021

    2021

  10. [18]

    Number 51/2021

    Natascha Hinterlang and Alina Tänzer.Optimal monetary policy using reinforcement learning. Number 51/2021. Deutsche Bundesbank Discussion Paper, 2021

    2021

  11. [19]

    Large language models as simulated economic agents: What can we learn from homo silicus? Technical report, National Bureau of Economic Research, 2023

    John J Horton. Large language models as simulated economic agents: What can we learn from homo silicus? Technical report, National Bureau of Economic Research, 2023

    2023

  12. [20]

    Michael Bradley Johanson, Edward Hughes, Finbarr Timbers, and Joel Z. Leibo. Emergent Bartering Behaviour in Multi-Agent Reinforcement Learning, May 2022. 11

    2022

  13. [21]

    An Economy of Neural Networks: Learning from Heterogeneous Experiences, October 2021

    Artem Kuriksha. An Economy of Neural Networks: Learning from Heterogeneous Experiences, October 2021

    2021

  14. [22]

    Econagent: Large language model-empowered agents for simulating macroeconomic activities

    Nian Li, Chen Gao, Mingyu Li, Yong Li, and Qingmin Liao. Econagent: Large language model-empowered agents for simulating macroeconomic activities. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 15523–15536, 2024

    2024

  15. [23]

    An empirical model of labor supply in a life-cycle setting.Journal of Political Economy, 89(6):1059–1085, 1981

    Thomas E MaCurdy. An empirical model of labor supply in a life-cycle setting.Journal of Political Economy, 89(6):1059–1085, 1981

    1981

  16. [24]

    Cengage Learning, 2021

    N Gregory Mankiw.Principles of Economics. Cengage Learning, 2021

    2021

  17. [25]

    Oxford university press New York, 1995

    Andreu Mas-Colell, Michael Dennis Whinston, Jerry R Green, et al.Microeconomic theory, volume 1. Oxford university press New York, 1995

    1995

  18. [26]

    Taxai: A dynamic economic simulator and benchmark for multi-agent reinforcement learning

    Qirui Mi, Siyu Xia, Yan Song, Haifeng Zhang, Shenghao Zhu, and Jun Wang. Taxai: A dynamic economic simulator and benchmark for multi-agent reinforcement learning. InProceedings of the 23rd International Conference on Autonomous Agents and Multiagent Systems, pages 1390–1399, 2024

    2024

  19. [27]

    Mf-llm: Simulating collective decision dynamics via a mean-field large language model framework.arXiv preprint arXiv:2504.21582, 2025

    Qirui Mi, Mengyue Yang, Xiangning Yu, Zhiyu Zhao, Cheng Deng, Bo An, Haifeng Zhang, Xu Chen, and Jun Wang. Mf-llm: Simulating collective decision dynamics via a mean-field large language model framework.arXiv preprint arXiv:2504.21582, 2025

    2025 arXiv

  20. [28]

    Learning macroeconomic poli- cies based on microfoundations: A stackelberg mean field game approach.arXiv preprint arXiv:2403.12093, 2024

    Qirui Mi, Zhiyu Zhao, Siyu Xia, Yan Song, Jun Wang, and Haifeng Zhang. Learning macroeconomic poli- cies based on microfoundations: A stackelberg mean field game approach.arXiv preprint arXiv:2403.12093, 2024

    2024 arXiv

  21. [29]

    Machine learning: an applied econometric approach.Journal of Economic Perspectives, 31(2):87–106, 2017

    Sendhil Mullainathan and Jann Spiess. Machine learning: an applied econometric approach.Journal of Economic Perspectives, 31(2):87–106, 2017

    2017

  22. [30]

    Michelle JK Osterman, Brady E Hamilton, Joyce A Martin, Anne K Driscoll, and Claudia P Valenzuela. Births: final data for 2022.National Vital Statistics Reports: From the Centers for Disease Control and Prevention, National Center for Health Statistics, National Vital Statisti...

    2022

  23. [31]

    Deep Reinforcement Learning for Optimal Investment and Saving Strategy Selection in Heterogeneous Profiles: Intelligent Agents working towards retirement, June 2022

    Fatih Ozhamaratli and Paolo Barucca. Deep Reinforcement Learning for Optimal Investment and Saving Strategy Selection in Heterogeneous Profiles: Intelligent Agents working towards retirement, June 2022

    2022

  24. [32]

    Generative agents: Interactive simulacra of human behavior

    Joon Sung Park, Joseph O’Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael S Bernstein. Generative agents: Interactive simulacra of human behavior. InProceedings of the 36th Annual ACM Symposium on User Interface Software and Technology, pages 1–22, 2023

    2023

  25. [33]

    Agentsociety: Large-scale simulation of llm-driven generative agents advances understanding of human behaviors and society.arXiv preprint arXiv:2502.08691, 2025

    Jinghua Piao, Yuwei Yan, Jun Zhang, Nian Li, Junbo Yan, Xiaochong Lan, Zhihong Lu, Zhiheng Zheng, Jing Yi Wang, Di Zhou, et al. Agentsociety: Large-scale simulation of llm-driven generative agents advances understanding of human behaviors and society.arXiv preprint arXiv:2502....

    2025 arXiv

  26. [34]

    Econojax: A fast & scalable economic simulation in jax.arXiv preprint arXiv:2410.22165, 2024

    Koen Ponse, Aske Plaat, Niki van Stein, and Thomas M Moerland. Econojax: A fast & scalable economic simulation in jax.arXiv preprint arXiv:2410.22165, 2024

    2024 arXiv

  27. [35]

    Society-in-the-loop: programming the algorithmic social contract.Ethics and Information Technology, 20(1):5–14, 2018

    Iyad Rahwan. Society-in-the-loop: programming the algorithmic social contract.Ethics and Information Technology, 20(1):5–14, 2018

    2018

  28. [36]

    Learning from zero: How to make consumption-saving decisions in a stochastic environment with an AI algorithm, February 2022

    Rui and Shi. Learning from zero: How to make consumption-saving decisions in a stochastic environment with an AI algorithm, February 2022

    2022

  29. [37]

    Using elasticities to derive optimal income tax rates.The Review of Economic Studies, 68(1):205–229, 2001

    Emmanuel Saez. Using elasticities to derive optimal income tax rates.The Review of Economic Studies, 68(1):205–229, 2001

    2001

  30. [38]

    Intelligence in the economy: Emergent behaviour in international trade modelling with reinforcement learning

    Abraham Ayooluwa Odukoya Sch. Intelligence in the economy: Emergent behaviour in international trade modelling with reinforcement learning. 2021

    2021

  31. [39]

    Can an AI agent hit a moving target.arXiv preprint arXiv, 2110, 2021

    Rui Aruhan Shi. Can an AI agent hit a moving target.arXiv preprint arXiv, 2110, 2021

    2021

  32. [40]

    MIT press Cambridge, 1998

    Richard S Sutton, Andrew G Barto, et al.Reinforcement learning: An introduction, volume 1. MIT press Cambridge, 1998

    1998

  33. [41]

    Discretion versus policy rules in practice

    John B Taylor. Discretion versus policy rules in practice. InCarnegie-Rochester conference series on public policy, volume 39, pages 195–214. Elsevier, 1993

    1993

  34. [42]

    Perspectives and challenges of agent-based simulation as a tool for economics and other social sciences.AAMAS, pages 35–42, 2009

    Klaus G Troitzsch. Perspectives and challenges of agent-based simulation as a tool for economics and other social sciences.AAMAS, pages 35–42, 2009. 12

    2009

  35. [43]

    Norton New York, 1992

    Hal R Varian and Hal R Varian.Microeconomic analysis, volume 3. Norton New York, 1992

    1992

  36. [44]

    Tradingagents: Multi-agents llm financial trading framework.arXiv preprint arXiv:2412.20138, 2024

    Yijia Xiao, Edward Sun, Di Luo, and Wei Wang. Tradingagents: Multi-agents llm financial trading framework.arXiv preprint arXiv:2412.20138, 2024

    2024 arXiv

  37. [45]

    Yangyang Yu, Zhiyuan Yao, Haohang Li, Zhiyang Deng, Yuechen Jiang, Yupeng Cao, Zhi Chen, Jordan Suchow, Zhenyu Cui, Rong Liu, et al. Fincon: A synthesized llm multi-agent system with conceptual verbal reinforcement for enhanced financial decision making.Advances in Neural Info...

    2024

  38. [46]

    The ai economist: Taxation policy design via two-level deep multiagent reinforcement learning.Science Advances, 8(18):eabk2607, 2022

    Stephan Zheng, Alexander Trott, Sunil Srinivasa, David C Parkes, and Richard Socher. The ai economist: Taxation policy design via two-level deep multiagent reinforcement learning.Science Advances, 8(18):eabk2607, 2022

    2022

  39. [47]

    Deep learning in economics: a systematic and critical review.Artificial Intelligence Review, 56(9):9497–9539, 2023

    Yuanhang Zheng, Zeshui Xu, and Anran Xiao. Deep learning in economics: a systematic and critical review.Artificial Intelligence Review, 56(9):9497–9539, 2023. 13 A Additional Experimental Results Table 6: Simulation results under varying numbers of government agents in EconGym...

    2023

  40. [48]

    Multi-government structures unlock richer coordination opportunities.Adding government agents expands the policy space and enables more inter-agent collaboration. In several double- government scenarios, such as ID-7 and ID-9 (GDP:2.45e+10, Welfare:1.21e+06), we observe notabl...

  41. [49]

    For instance, ID-21 (LLM-only) results in GDP:1.98e+09and Welfare:8.42e+04; while ID-22 (PPO-only) terminates with GDP:1.50e+09, Welfare:-8.91e+04

    Pure AI policies often underperform in complex economic settings.AI-only agents frequently struggle to find effective strategies. For instance, ID-21 (LLM-only) results in GDP:1.98e+09and Welfare:8.42e+04; while ID-22 (PPO-only) terminates with GDP:1.50e+09, Welfare:-8.91e+04....

  42. [50]

    For example, ID-28 14 (Saez tax + DDPG + LLM) achieves the highest GDP:3.06e+10, strong Welfare:1.17e+06, and the lowest Gini:0.32among all tested setups

    Hybrid approaches (AI + economics) consistently deliver superior results.Top-performing configurations integrate structured economic rules with adaptive AI agents. For example, ID-28 14 (Saez tax + DDPG + LLM) achieves the highest GDP:3.06e+10, strong Welfare:1.17e+06, and the...

    2022

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.