pith. sign in

arxiv: 2606.26883 · v1 · pith:GXRUE47Nnew · submitted 2026-06-25 · 💻 cs.DL

EconSimulacra: A Digital Twin Platform of Socio-Economic Systems Powered by LLM Agents

Ryuji Hashimoto , Masahiro Kaneko , Kentaro Ueda , Takehiro Takayanagi , Kiyoshi Izumi This is my paper

Pith reviewed 2026-06-26 01:52 UTC · model grok-4.3

classification 💻 cs.DL
keywords LLM agentsmulti-agent simulationsocio-economic modelingdigital twincross-domain dynamicsinternal state mechanismsocial networksconsumer behavior
0
0 comments X

The pith

Shared internal states let LLM agents produce realistic links between online attention and offline popularity in one simulator.

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

The paper introduces a multi-agent simulator in which agents draw from experiences in consumer economy, mobility, and social networks. These experiences are stored in memory and converted into a single shared internal state such as stress level. The state then shapes decisions in any domain, letting feedback loops arise directly from individual choices. A case study demonstrates that this mechanism generates a nonlinear relationship between online social attention and offline local popularity. The approach aims to create unified artificial societies instead of modeling domains in isolation.

Core claim

EconSimulacra couples consumer economy, mobility, and social networks through a shared internal-state mechanism. Experiences accumulated across domains are stored in memory and transformed into shared internal states such as stress level. This connects heterogeneous domains through individual decision making and allows agents to reconcile competing demands. The design reproduces a nonlinear relationship between online social attention and offline local popularity, showing how realistic cross-domain dynamics can emerge within a unified artificial society.

What carries the argument

The shared internal-state mechanism that converts accumulated memory from multiple domains into a scalar value such as stress level to guide coherent agent decisions across those domains.

If this is right

  • Agents can reconcile demands from economy, mobility, and social networks using one internal value.
  • Nonlinear cross-domain dynamics such as attention-to-popularity links appear without separate models for each domain.
  • The platform functions as a digital twin for exploring how changes in one domain affect others through individual choices.
  • Unified artificial societies become feasible for studying feedback between online activity and offline behavior.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Extending the memory-to-state conversion to include time-decay or event-specific weighting could allow agents to respond differently to recent versus distant experiences.
  • Tracking how internal-state distributions evolve across agent populations might reveal emergent patterns of inequality or segregation in simulated societies.
  • Connecting the simulator to live data streams would test whether the same mechanism sustains realistic dynamics under changing external conditions.

Load-bearing premise

LLM-generated agent decisions driven only by memory turned into a scalar internal state produce coherent and realistic cross-domain behavior without additional domain-specific rules or external calibration.

What would settle it

A direct comparison of the nonlinear curve between simulated online attention and offline local popularity against the same relationship measured in real social-media and sales data from actual locations.

Figures

Figures reproduced from arXiv: 2606.26883 by Kentaro Ueda, Kiyoshi Izumi, Masahiro Kaneko, Ryuji Hashimoto, Takehiro Takayanagi.

Figure 1
Figure 1. Figure 1: Overview of EconSimulacra. Agents continuously accumulate heterogeneous experiences from the [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Structure of the JSON-based configuration [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Simplified execution flow of a simulation. At [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: , the interface allows users to inspect the states of individual domains, such as the grid-based environment, the social network, and macroeco￾nomic indicators, while simultaneously monitoring the internal states and decision-making processes of individual agents. Users can interactively ex￾amine agent attributes, memories, stress levels, in￾ventories, and action histories, enabling a detailed understandin… view at source ↗
Figure 5
Figure 5. Figure 5: Time series of sales amounts for Pizza Place [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Behavioral clusters derived from normalized daytime consumption vectors and their temporal evolution in [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Time series of the topic composition of online [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗
read the original abstract

Real-world social behavior emerges from tightly coupled domains: economic conditions shape mobility and social interactions, while online attention and offline activity feed back into local popularity and consumer behavior. Capturing these feedback loops requires artificial societies in which agents carry experiences from one domain into decisions in another. Large language models (LLMs) provide a promising foundation for such societies. However, existing LLM-based simulators typically model domains in isolation or merely place them side by side. To enable such cross-domain interactions, we present EconSimulacra, a multi-agent social simulator that couples consumer economy, mobility, and social networks through a shared internal-state mechanism. In EconSimulacra, experiences accumulated across different domains are stored in memory and transformed into shared internal states (i.e., stress level) connecting heterogeneous domains through individual decision making. This design allows agents to reconcile competing demands arising from multiple domains and generate coherent cross-domain behaviors. As a case study, we show that the shared internal state mechanisms reproduce a nonlinear relationship between online social attention and offline local popularity, illustrating how realistic cross-domain dynamics can emerge within a unified artificial society.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper introduces EconSimulacra, a multi-agent LLM-based simulator that couples consumer economy, mobility, and social networks through a shared internal-state mechanism in which cross-domain memories are aggregated into a scalar (stress level) that informs agent decisions. The central claim is that this design enables coherent cross-domain behaviors, demonstrated via a case study in which the mechanism reproduces a nonlinear relationship between online social attention and offline local popularity.

Significance. If the case-study result holds under external validation, the shared-state approach would address a recognized limitation of existing LLM simulators (isolated or side-by-side domains) and supply a concrete mechanism for emergent cross-domain feedback in artificial societies. The absence of any reported parameter count, ablation, or external benchmark data, however, leaves the practical significance of the platform undetermined.

major comments (2)
  1. [Abstract / Case Study] Abstract / Case Study section: the claim that the shared internal state 'reproduces a nonlinear relationship' is presented without any quantitative metrics, simulation scale (agent count, time steps), memory-to-state mapping equations, or comparison to real-world distributions. This omission renders the central demonstration impossible to evaluate for independence from the generative loop.
  2. [Shared internal-state mechanism] Shared internal-state mechanism (described in the abstract and presumably §3–4): aggregation of heterogeneous memories (consumer, mobility, social) into a single scalar stress level is asserted to enable agents to 'reconcile competing demands,' yet no ablation that removes the scalar or substitutes domain-specific states is reported. Without such a test, it is unclear whether the scalar is information-preserving or whether the observed nonlinearity arises from LLM priors rather than the claimed mechanism.
minor comments (1)
  1. The abstract would be strengthened by a single sentence stating the LLM model, number of agents, and any external data sources used for the case study.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive report. The comments highlight important aspects of clarity and validation that will strengthen the manuscript. We address each major comment below and indicate the revisions we will make.

read point-by-point responses

  1. Referee: [Abstract / Case Study] Abstract / Case Study section: the claim that the shared internal state 'reproduces a nonlinear relationship' is presented without any quantitative metrics, simulation scale (agent count, time steps), memory-to-state mapping equations, or comparison to real-world distributions. This omission renders the central demonstration impossible to evaluate for independence from the generative loop.

    Authors: We agree that the case-study demonstration requires more quantitative support to allow independent evaluation. In the revised manuscript we will expand the case-study section (and update the abstract accordingly) to report: (i) the precise simulation scale (number of agents and time steps), (ii) the explicit memory-to-state aggregation equations that map heterogeneous memories to the scalar stress level, (iii) quantitative metrics (e.g., correlation or regression coefficients) characterizing the reproduced nonlinear relationship, and (iv) a direct comparison of the simulated distribution against available real-world data on online attention versus offline popularity. These additions will make the independence of the result from the generative loop verifiable. revision: yes

  2. Referee: [Shared internal-state mechanism] Shared internal-state mechanism (described in the abstract and presumably §3–4): aggregation of heterogeneous memories (consumer, mobility, social) into a single scalar stress level is asserted to enable agents to 'reconcile competing demands,' yet no ablation that removes the scalar or substitutes domain-specific states is reported. Without such a test, it is unclear whether the scalar is information-preserving or whether the observed nonlinearity arises from LLM priors rather than the claimed mechanism.

    Authors: We concur that an ablation study is necessary to isolate the contribution of the shared scalar. In the revision we will add an ablation experiment in which the single stress-level scalar is replaced by independent domain-specific internal states. We will then compare the resulting cross-domain behaviors (specifically the nonlinear attention-popularity relationship) against the original shared-state configuration. This will provide direct evidence that the observed nonlinearity is attributable to the cross-domain aggregation mechanism rather than to LLM priors alone. revision: yes

Circularity Check

0 steps flagged

No circularity: emergent simulation outcome independent of input reduction

full rationale

The paper introduces a multi-agent LLM simulator with shared internal states (stress level) derived from cross-domain memory to enable coherent agent decisions across economy, mobility, and social networks. The reported nonlinear online-offline relationship is presented as an emergent illustration from running the simulator, not as a mathematical derivation, fitted parameter, or self-citation-dependent uniqueness theorem. No equations, ansatzes, or load-bearing self-citations appear in the provided text that would reduce the claimed reproduction to the design inputs by construction. The platform is self-contained as a generative model whose outputs are not forced to match any withheld external benchmark via the mechanisms described.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The platform rests on the untested premise that LLM memory plus a scalar internal state is sufficient to mediate realistic trade-offs across domains; no free parameters, axioms, or invented entities are enumerated in the abstract.

pith-pipeline@v0.9.1-grok · 5747 in / 1147 out tokens · 18571 ms · 2026-06-26T01:52:54.186819+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

17 extracted references · 4 linked inside Pith

  1. [1]

    InFindings of the Associa- tion for Computational Linguistics: EMNLP 2020, pages 1644–1650

    TweetE- val: Unified benchmark and comparative evaluation for tweet classification. InFindings of the Associa- tion for Computational Linguistics: EMNLP 2020, pages 1644–1650. Nicolas Bougie and Narimawa Watanabe

    2020

  2. [2]

    In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing: Industry Track, pages 215–229

    CitySim: Modeling urban behaviors and city dynam- ics with large-scale LLM-driven agent simulation. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing: Industry Track, pages 215–229. Ayush Chopra, Shashank Kumar, Nurullah Giray Kuru, Ramesh Raskar, and Arnau Quera-Bofarull

    2025

  3. [3]

    InFindings of the Association for Com- putational Linguistics: NAACL 2024, pages 3326–

    Simulating opinion dynamics with networks of LLM- based agents. InFindings of the Association for Com- putational Linguistics: NAACL 2024, pages 3326–

    2024

  4. [4]

    Epstein and Robert L

    Joshua M. Epstein and Robert L. Axtell. 1996.Growing artificial societies: Social science from the bottom up. The MIT Press. Chen Gao, Xiaochong Lan, Nian Li, Yuan Yuan, Jingtao Ding, Zhilun Zhou, Fengli Xu, and Yong Li

    1996

  5. [5]

    Preprint, arXiv:2203.05794

    BERTopic: Neural topic modeling with a class-based TF-IDF procedure. Preprint, arXiv:2203.05794. Torsten Hägerstrand

    Pith/arXiv arXiv

  6. [6]

    Ryuji Hashimoto, Masahiro Kaneko, Ryosuke Takata, Takehiro Takayanagi, and Kiyoshi Izumi

    What about people in re- gional science?Papers of the Regional Science Asso- ciation, 24(1):6–21. Ryuji Hashimoto, Masahiro Kaneko, Ryosuke Takata, Takehiro Takayanagi, and Kiyoshi Izumi. 2026a. From heard to lived opinions: Simulating opinion dynamics with grounded llm agents in economic en- vironments.Preprint, arXiv:2603.26701. Ryuji Hashimoto, Takehir...

    arXiv 2025

  7. [7]

    Building LLM-based artificial market simulations: Can LLMs function as agents in multi-agent simulations for finance? InThe Inter- national Conference on Principles and Practice of Multi-Agent Systems (PRIMA 2025), pages 56–71. C. S. Holling

    2025

  8. [8]

    Qun Ma, Xiao Xue, Deyu Zhou, Xiangning Yu, Donghua Liu, Xuwen Zhang, Zihan Zhao, Yifan Shen, Peilin Ji, Juanjuan Li, Gang Wang, and Wan- peng Ma

    AgentSims: An open-source sandbox for large language model evalu- ation.Preprint, arXiv:2308.04026. Qun Ma, Xiao Xue, Deyu Zhou, Xiangning Yu, Donghua Liu, Xuwen Zhang, Zihan Zhao, Yifan Shen, Peilin Ji, Juanjuan Li, Gang Wang, and Wan- peng Ma

    arXiv

  9. [9]

    Tianmi Ma, Jiawei Du, Wenxin Huang, Wenjie Wang, Liang Xie, Xian Zhong, and Joey Tianyi Zhou

    Computational experiments meet large language model based agents: A survey and perspective.Preprint, arXiv:2402.00262. Tianmi Ma, Jiawei Du, Wenxin Huang, Wenjie Wang, Liang Xie, Xian Zhong, and Joey Tianyi Zhou

    arXiv

  10. [10]

    InFindings of the Association for Computational Linguistics: EMNLP 2025, pages 5496–5514

    Agent trading arena: A study on numerical under- standing in LLM-based agents. InFindings of the Association for Computational Linguistics: EMNLP 2025, pages 5496–5514. Meta

    2025

  11. [11]

    Xinyi Mou, Xuanwen Ding, Qi He, Liang Wang, Jing- cong Liang, Xinnong Zhang, Libo Sun, Jiayu Lin, Jie Zhou, Huang Xuanjing, and Zhongyu Wei

    The Llama 3 herd of models.Preprint, arXiv:2407.21783. Xinyi Mou, Xuanwen Ding, Qi He, Liang Wang, Jing- cong Liang, Xinnong Zhang, Libo Sun, Jiayu Lin, Jie Zhou, Huang Xuanjing, and Zhongyu Wei

    Pith/arXiv arXiv

  12. [12]

    InFindings of the Association for Computational Linguistics: ACL 2024, pages 4789–4809

    Unveiling the truth and facilitating change: Towards agent-based large-scale social movement simulation. InFindings of the Association for Computational Linguistics: ACL 2024, pages 4789–4809. Önder Gürcan

    2024

  13. [13]

    InHybrid Human AI Systems for the Social Good (HHAI 2024), Frontiers in Artificial In- telligence and Applications

    LLM-augmented agent-based modelling for social simulations: Challenges and op- portunities. InHybrid Human AI Systems for the Social Good (HHAI 2024), Frontiers in Artificial In- telligence and Applications. IOS Press. OpenAI

    2024

  14. [14]

    Preprint, arXiv:2508.10925

    gpt-oss-120b & gpt-oss-20b model card. Preprint, arXiv:2508.10925. John F. Padgett and Walter W. Powell. 2012.The emer- gence of organizations and markets. Princeton Uni- versity Press. Joon Sung Park, Joseph O’Brien, Carrie Jun Cai, Mered- ith Ringel Morris, Percy Liang, and Michael S. Bern- stein

    Pith/arXiv arXiv 2012

  15. [15]

    Alexander Sasha Vezhnevets, John P

    Qwen3 technical report.Preprint, arXiv:2505.09388. Alexander Sasha Vezhnevets, John P. Agapiou, Avia Aharon, Ron Ziv, Jayd Matyas, Edgar A. Duéñez- Guzmán, William A. Cunningham, Simon Osindero, Danny Karmon, and Joel Z. Leibo

    Pith/arXiv arXiv

  16. [16]

    Preprint, arXiv:2312.03664

    Gener- ative agent-based modeling with actions grounded in physical, social, or digital space using Concordia. Preprint, arXiv:2312.03664. Jun Zhang, Yuwei Yan, Junbo Yan, Zhiheng Zheng, Jinghua Piao, Depeng Jin, and Yong Li

    arXiv

  17. [17]

    festive / discounted / christmaseve

    increases markedly, while the relative sizes of the routine shopping clusters decrease. These observa- tions suggest that the promotional event reshapes the distribution of consumption behaviors across the population. Emergence of Online DiscussionsTo investi- gate how changes in offline activities are reflected in the social network domain, we extract di...

    2022