EconAI: Dynamic Persona Evolution and Memory-Aware Agents in Evolving Economic Environments

Society, culture, and collective behavior2026arXivApproved editorial review

Authors: Annie Liu, Zane Cao, Lang Chen, Zongxin Xu, Zigan Wang

Keywords: Economic agent simulation, Long-term and short-term memory, Economic Sentiment Index, Dynamic persona claim, Macroeconomic simulation, Construct validity, Reproducibility, Publication provenance

Source: Open primary source (opens in a new tab)

5
Authors
13
Findings
15
Limitations
4
Evidence

Editorial summary

English

EconAI proposes a GPT-4o-mini-driven agent-based economy. Households choose work and consumption; firms produce, invest, and hire; government and finance provide taxation, redistribution, and interest. The LLM layer is combined with explicit economic rules: Cobb-Douglas production, capital, demand, random price and wage adjustments, and a stated Taylor-rule mechanism. Long-term memory stores embedded event summaries, short-term memory retains current context, and an LLM-generated Economic Sentiment Index is smoothed over time to modulate work and consumption.

The title promises dynamic persona evolution, but each experimental persona is only a name, age, and occupation string reinjected verbatim every month. The prose claims persona extraction and continuous updates, yet supplies no prompt, representation, update rule, longitudinal example, ablation, or persona outcome. The evaluated construct is stateful economic behavior with memory and sentiment, not personality change or psychological validity.

The evaluation plots 20 simulated years of inflation, nominal GDP, growth, and unemployment against LEN, CATS, Composite, AI-Eco/AI-Econ, and EconAgent. EconAI is generally visually smoother. The paper also reports a Phillips relationship with Pearson rho=-0.522 and p<0.01, an Okun plot, two firm traces interpreted as competition and cooperation, a ten-year ablation, a textual COVID-19 shock, and inflation at 100, 200, and 300 agents.

No panel compares simulation with observed economic series or reports a loss, error, predictive target, or accepted plausibility range. There are no seeds, repeats, error bars, intervals, raw data, or notebooks. The figures can support visual stability in one run, not greater accuracy, precision, or replication of real cycles. The text says four baselines but names five, and says AI-Eco is dropped from macro plots even though all four legends include an AI-Econ series. Phillips omits Composite and AI-Eco; Okun omits Composite.

The ablation removes history, sentiment, belief, or investment, not persona, and consists of five ten-point lines without repeats or uncertainty. The COVID prompt may elicit pretrained knowledge of the pandemic rather than validate a causal economic mechanism; its figure includes Normal, EconAgent, and EconAI instead of a clearly labeled treatment/counterfactual pair. The scale check shows only inflation and defines no equivalence threshold.

The specification is also incomplete. The capital equation places K_t on both sides; exact lambda, ESI scale, theta, beta, confidence construction, prompts, parser, temperature, GPT-4o-mini snapshot, seed, and retry policy are missing. The event summarizer is called instruction-tuned without a base model, data, split, checkpoint, or evaluation. The manuscript twice points to an Appendix and to supplementary statistics that are absent from the arXiv bundle. No safely attributable repository, code, or data was located for this version.

There is a serious provenance conflict. The official MALGAI 2026 site lists the title as an oral and links OpenReview ibfvPld90A, but that record assigns the same title to Yijin Chen, Ning Lyu, Shengning Lang, Hao Yan, Zhiguo Tao, Xiaotong Ding, and Xiaotong Zhu rather than the five arXiv authors; its indexed PDF shares the abstract and substantial text. The workshop occurred on April 27 and the arXiv record appeared on May 13. A withdrawn ICLR 2025 EconAI submission with a different title and related author names also exists. Without an official resolution, the workshop oral is not attributed here to the arXiv author list.

Overall, this is an interesting LLM economic-simulation proposal with memory and sentiment, supported by qualitative illustrations. It does not establish personality evolution, human likeness, empirical realism, macroeconomic precision, causality, or reproducibility, and should not be treated as evidence for economic policy or other high-stakes decisions.

Español

EconAI propone una economía agent-based impulsada por GPT-4o-mini. Los hogares deciden si trabajan y qué proporción consumen; las empresas producen, invierten y contratan; gobierno y sistema financiero aplican impuestos, redistribución e interés. La capa LLM se combina con reglas económicas explícitas: producción Cobb-Douglas, actualización de capital, demanda, ajustes aleatorios de precios y salarios y una regla de Taylor anunciada. Una memoria larga guarda resúmenes vectorizados, la corta conserva contexto reciente y un Economic Sentiment Index generado por el LLM se suaviza en el tiempo para modular trabajo y consumo.

El título promete evolución dinámica de persona, pero el experimento solo define cada persona como nombre, edad y ocupación, y vuelve a inyectar esa cadena literalmente cada mes. El texto afirma que extrae y actualiza personas, pero no publica prompt, representación, regla de actualización, ejemplo longitudinal, ablación ni métrica de personalidad. Lo que sí se prueba es estado económico con memoria y sentimiento; no cambio de personalidad ni validez psicológica.

La evaluación muestra 20 años simulados de inflación, PIB nominal, crecimiento y desempleo frente a LEN, CATS, Composite, AI-Eco/AI-Econ y EconAgent. EconAI suele dibujar curvas más suaves. También presenta una relación de Phillips con rho de Pearson -0,522 y p<0,01, una figura de Okun, dos trayectorias de empresas interpretadas como competencia/cooperación, una ablación de diez años, un shock textual de COVID-19 y una comparación de inflación con 100, 200 y 300 agentes.

Estas figuras no comparan con series económicas observadas ni aportan pérdida, error, intervalo de plausibilidad o ajuste predictivo. No hay seeds, réplicas, barras de error, intervalos, datos brutos o notebooks. Por tanto, sostienen como máximo que una ejecución es visualmente más estable; no sostienen mayor precisión, exactitud o mejor reproducción de ciclos reales. El texto dice cuatro baselines pero enumera cinco. Además afirma que AI-Eco se excluye de las gráficas macro por degenerar, aunque las cuatro leyendas contienen una serie AI-Econ. Phillips omite Composite y AI-Eco; Okun omite Composite.

La ablación elimina historia, sentimiento, belief o inversión, pero no persona. Son cinco líneas de diez puntos sin repeticiones ni incertidumbre. El shock COVID inyecta la frase de emergencia nacional en el prompt: la respuesta puede proceder del conocimiento previo del modelo sobre COVID, no de un mecanismo causal validado. La figura incluye Normal, EconAgent y EconAI en vez de una pareja claramente etiquetada treatment/counterfactual. La robustez por escala solo muestra inflación y no define equivalencia.

La especificación tampoco permite reconstrucción. La ecuación de capital coloca K_t a ambos lados; faltan lambda exacta, escala ESI, theta, beta, cálculo de confianza, prompts, parser, temperatura, snapshot de GPT-4o-mini, seed y protocolo de retry. El event summarizer se declara instruction-tuned sin modelo base, datos, split, checkpoint o evaluación. El paper remite dos veces a un Appendix y a estadísticas suplementarias que no existen en el paquete arXiv. No hay repo, datos ni código atribuibles con seguridad a esta versión.

Existe además un conflicto grave de procedencia. La web oficial de MALGAI 2026 lista el título como oral y enlaza OpenReview ibfvPld90A, pero ese registro atribuye el mismo título a Yijin Chen, Ning Lyu, Shengning Lang, Hao Yan, Zhiguo Tao, Xiaotong Ding y Xiaotong Zhu, no a los cinco autores del arXiv, y su PDF indexado comparte abstract y texto sustancial. El workshop fue el 27 de abril y el arXiv apareció el 13 de mayo. También hay un EconAI retirado de ICLR 2025 con otro título y autores relacionados. Sin una aclaración oficial, no atribuimos la aceptación oral a los autores del arXiv.

En conjunto, el trabajo es una propuesta interesante de simulador económico LLM con memoria y sentimiento, apoyada por ilustraciones cualitativas. No demuestra evolución de personalidad, humanidad, realismo empírico, precisión macroeconómica, causalidad o reproducibilidad. Tampoco debe usarse como evidencia para política económica o decisiones de alto riesgo.

Research question

Can an agent-based economy with long/short memory, LLM-generated economic sentiment, and labor-consumption decisions produce more stable macro trajectories and react to textual shocks? And does that demonstrate persona evolution?

Method

Monthly simulation with households, firms, government, and finance. GPT-4o-mini decides labor and consumption based on income, prices, savings, unemployment, interest, and ESI; explicit rules update production, capital, demand, prices, wages, taxes, and interest. The audit visually reviewed 10 pages, complete TeX, equations, nine panels, publication metadata, and artifact availability.

Sample: The main configuration appears to use 100 synthetic households and 20 years; the scale comparison uses 100, 200, and 300 agents, and the ablation ten years. The paper does not report number of seeds, repetitions, calls, failures, cost, exact sample for Pearson, or number of independent trajectories.

Findings

  • EconAI is visually smoother across several simulated trajectories.
  • The paper reports Phillips rho=-0.522 and p<0.01 without data or exact n.
  • Okun is asserted without coefficient or p-value.
  • There are no real series or fit metric for accuracy.
  • Four baselines are declared but five are listed.
  • AI-Eco is said to be excluded although AI-Econ appears in the macro legends.
  • The ablation does not evaluate persona extraction.
  • The COVID shock may exploit prior LLM knowledge.
  • The experimental persona is name, age, and occupation reinjected.
  • Personality evolution is not operationalized.
  • Promised appendix and supplement are not on arXiv.
  • There is no code or data attributable to the preprint.
  • The venue record has different authors.

Limitations

  • arXiv v1 preprint.
  • Unresolved conflict of authorship and venue.
  • No validation of personality or humanity.
  • No observed economic data in the evaluation.
  • No seeds, replicas, error bars, or intervals.
  • No raw data, code, or notebooks.
  • No prompts, parser, or generation parameters.
  • GPT-4o-mini model without exact snapshot.
  • Event summarizer without data or checkpoint.
  • Incomplete equations and parameters.
  • Absent appendix and supplementary statistics.
  • Inconsistent baseline accounting and labels.
  • Single-trajectory ablation without uncertainty.
  • COVID intervention confuses prompt with prior knowledge.
  • Scale robustness shown only for inflation.

What the study does not establish

  • Dynamic evolution of personality.
  • Valid psychological personality or human preferences.
  • Greater accuracy against real economic data.
  • Macroeconomic prediction.
  • Causal mechanism of the COVID shock.
  • Independent discovery of economic laws.
  • Causal effect of memory, sentiment, or investment.
  • Robustness across seeds or closed models.
  • Reproducibility of a figure.
  • MALGAI acceptance for the arXiv authors.
  • Suitability for policy or high-risk decisions.

Traceability

Scope: Full text

Version: arXiv:2605.13762v1, submitted 2026-05-13, 10 pages, complete TeX; unresolved authorship conflict with OpenReview ibfvPld90A

Consulted source: https://arxiv.org/abs/2605.13762v1

Review: Codex 10-page visual full-text, complete TeX, persona-construct, economic-metric, formula, publication-provenance and artifact-reproducibility audit, 2026-07-17

Approval: Codex fidelity pass, 2026-07-17

English translation: approved, 2026-07-18

Models evaluated

  • GPT-4o-mini, exact snapshot not reported
  • MiniLM proposed as an example text encoder, exact checkpoint not reported
  • Unspecified instruction-tuned event summarizer
  • LEN
  • CATS
  • Composite LEN/CATS
  • AI-Eco or AI-Econ, naming unresolved
  • EconAgent

Instruments and metrics

  • Long-term and short-term event memory
  • Economic Sentiment Index
  • LLM work and consumption decision prompt
  • Hand-designed macroeconomic environment equations
  • Phillips and Okun scatter plots
  • Ten-year component ablation
  • Textual COVID-19 intervention
  • 100/200/300-agent inflation comparison
  • Independent construct, metric, formula, provenance and artifact audit

Data used

  • 2018 U.S. Census adult age distribution, exact extract absent
  • Pareto wage distribution calibrated to 2018 U.S. tax filings, parameters absent
  • Simulated 20-year macroeconomic trajectories, raw data absent
  • Simulated Phillips and Okun points, raw data absent
  • Ten-year ablation traces, raw data absent
  • COVID-19 intervention traces, raw prompts and reflections absent

Evidence and location

  • Text, method, equations, figures, and limitations: arXiv:2605.13762v1; PDF sha256 eed632b2bf0c4992e1b2c0332af9d7911fe7140d9ec4d122eddab1e22daaa2a0; TeX sha256 416b37d6390254cd5778c24245fadd7b4ef52153a1ab832e98c2fbecd0b44d80
  • Official oral listing and record with different authors: https://iclr26-mal-gai.github.io/papers.html -> OpenReview ibfvPld90A; OpenReview MALGAI submissions published 2026-03-02
  • Prior withdrawn EconAI record: OpenReview HzG3A0VD1k, withdrawn ICLR 2025 submission
  • Independent audit: reports/verification/article-332-econai-persona-construct-publication-provenance-experiment-metric-artifact-and-reproducibility-audit.json