PaperlayerView exact revision on arXiv

Source-linked AI summary

From Economic Agents to Agentic Economies: A Systems Blueprint for Economic World Models

Jiale Han, Xiang Li, Jing Qian, Wenyuan Gu, Pin Gao, Ye Luo, Hongyuan Zha, Dacheng Tao, Benyou Wang, Lin William Cong

arXiv:2608.06020v1cs.AIcs.LG

TL;DR

Economic research often explains observed outcomes without showing how they emerge from interacting agents, markets, and institutions. This paper proposes an implementation roadmap and six-level capability ladder for Economic World Models, finding that 737 validated papers remain concentrated in lower-level agent systems while evolving worlds and real-world alignment are rare.

  • Problem

    Economic research needs simulators that generate aggregate outcomes from interacting agents, markets, and institutions rather than only observing or explaining those outcomes.

  • Method

    The paper develops an implementation roadmap, six-level capability ladder, systematic literature survey, and agent–environment framework for Economic World Models.

  • Results

    737 validated EWM papers reveal research concentrated in L1–L3 agent intelligence, while self-evolving agents, evolving institutions, and sim-to-real alignment remain rare.

  • Takeaways & Limitations

    EWMs are positioned as programmable sandboxes for testing policies, market designs, firm strategies, and AI-agent training, planning, evaluation, and safety.

  • Takeaways & Limitations

    The field has not yet integrated self-evolving agents, evolving institutions, and online real-world alignment within one economic world.

Abstract

from arXiv · show

Economic World Models (EWMs) are generative economic models that simulate how economies evolve from within by modeling heterogeneous agents, their beliefs and actions, and the market and institutional mechanisms through which their interactions produce aggregate outcomes. This paper develops an implementation roadmap for building economic world models as generative engines in which heterogeneous agents act, interact, adapt, and co-evolve with markets and institutions, thereby producing economic dynamics from the inside. We organize EWM systems into a six-level capability ladder, from fixed rule-based agent worlds to adaptive and LLM-based agent worlds, self-evolving agents, evolving institutional worlds, and sim-to-real economic twins aligned with real observations. A systematic literature survey across these levels reveals that existing work remains concentrated in lower-level agent and simulation environments, while systems with self-evolving agents, endogenous institutions, persistent empirical alignment, and validated economic mechanisms remain rare. By translating the EWM agenda into an implementation blueprint, this paper aims to accelerate the development of the next generation of economic simulation environments that can serve as high-fidelity sandboxes for human decision-makers and as training, planning, evaluation, and safety substrates for AI agents. We release a curated paper list and related resources to support future research.

1 Introduction

Economic World Models extend economic modeling from observing and explaining aggregate outcomes to generatively simulating how heterogeneous agents act, interact, and adapt within markets and institutions. The paper frames EWMs as modular runtimes that create executable feedback loops for agentic economies and support decision-making, AI training, planning, and evaluation.

  • Economic World Models: EWMs model economic state transitions as outcomes generated by heterogeneous agents interacting through markets, institutions, and evolving rules.Unlike physical worlds governed by external laws, economic dynamics are generated from within the system.
  • Economic World Models: An EWM is a generative engine that predicts an economy’s evolution by modeling how agents observe, reason, act, interact, and adapt under institutions and constraints.Its core runtime is an action–state loop: agent actions produce a next state that reshapes later beliefs, constraints, memories, and strategies.
  • The goal: towards agentic economies: Agentic Economies add an executable economic feedback loop in which actions’ consequences are realized, evaluated, and returned as learning signals, moving beyond isolated next-action generation.This enables iterative co-improvement between Economic Agents and the EWM.
  • Implementation: The implementation uses a modular economic runtime coupling agent, environment, co-evolution, and institutional layers rather than a monolithic predictor.The agent layer represents heterogeneous actors, while the environment layer encodes states, market mechanisms, contracts, accounting identities, and institutional rules.
  • Applications of EWMs: EWMs can serve human decision-makers as high-fidelity sandboxes and AI systems as training, planning, and evaluation substrates under scarcity, incentives, and institutional constraints.Applications include testing policies, market designs, firm strategies, systemic risks, and institutional reforms before deployment.

2 Economic World Models as AI Systems

Economic World Models represent economies as computable dynamical systems in which heterogeneous, subjective agents interact through endogenous markets and institutions. Their implementation requires state representations and transitions that capture beliefs, actions, adaptation, evolving rules, and alignment with the real economy.

  • Economic World Models as AI Systems: Economic worlds differ from physical worlds because subjective agents form beliefs, make strategic decisions, and revise behavior under incomplete information and institutional constraints.Unlike physical objects governed primarily by objective laws, households, firms, banks, investors, and regulators interpret information before acting.
  • Economic World Models as AI Systems: Prices, allocations, quantities, risks, and macroeconomic conditions emerge from decentralized trading, bargaining, bidding, contracting, and market clearing.These outcomes are endogenous rather than externally imposed, and they feed back into subsequent agent decisions.
  • Economic World Models as AI Systems: An economic state combines aggregate conditions, agent-side private and belief states, and the institutional environment governing agent behavior.It includes variables such as prices, employment, wealth, inventories, risk exposure, expectations, perceived risks, policies, information channels, and market mechanisms.
  • Economic World Models as AI Systems: EWM transitions update agent beliefs, predict actions, update individual states, predict the next economic state, and evolve rules and mechanisms.The belief-update operator incorporates prior beliefs, current and aggregate states, interventions, and institutions; decision rules then map updated beliefs and agent states to actions.
  • Economic World Models as AI Systems: A minimal EWM execution interface constructs a named world, resets its state, collects parallel agent actions, and advances the economy with the joint action dictionary.The environment-style rollout uses heterogeneous populations such as households, firms, and banks, while advanced hooks include logging, evolution, and real-data alignment.
  • Economic World Models as AI Systems: Strong EWM systems require endogenous closure, behavioral fidelity, evolving dynamics, and reality alignment.Agents should perceive incentives, information, and constraints; strategies, cognition, and governing rules should evolve; and the system should remain aligned with the real economy over time.

3 A Systematic Empirical Study of the EWM Literature Landscape

The paper organizes Economic World Models into six progressively stronger capability levels and systematically classifies validated literature using this taxonomy. From 7,836 candidate papers, 737 validated papers reveal rapid progress concentrated in agent capabilities, while self-evolving agents, evolving institutions, and sim-to-real alignment remain rare.

  • Capability taxonomy: Six capability levels operationalize EWM engineering desiderata, progressing from fixed rule-based agents to adaptive agents, LLM-based agents, evolving agents, co-evolving institutions, and sim-to-real twins.The hierarchy is an implementation taxonomy rather than an equilibrium concept, and higher capability does not guarantee counterfactual consistency.
  • Survey methodology: The survey classifies each validated paper using the six-level coding rules defined in Table 2.The coding framework is applied systematically across the existing literature.
  • Empirical landscape: 7,836 candidate papers yielded 737 validated EWM papers, whose distribution was summarized across years, source types, and capability levels.The resulting corpus provides the empirical basis for analyzing the field’s development and capability concentration.
  • Empirical landscape: Recent EWM research has accelerated sharply around 2024–2025, with most growth concentrated in L1–L3 adaptive rule-based and LLM-based autonomous-agent systems.Before the early 2010s, annual output was small and dominated by L1–L2 systems; volume grew steadily after 2018.
  • Research frontier: Systems with self-evolving agents, endogenous institutional evolution, and L6 sim-to-real correction remain rare, leaving most simulations with fixed economic environments.L6 systems are nearly absent, with only limited evidence of repeated correction against real-world observations.

4 Implementation Architecture for EWM Systems

EWM systems are organized as modular economic actors, executable environments, endogenous co-evolution, and empirical alignment, with reusable interfaces governing both configuration and runtime execution. This architecture supports adaptive economic worlds rather than static simulations while preserving correction against observed behavior.

  • Core architecture: EWM implementations separate economic actors, executable worlds, endogenous co-evolution, and empirical alignment while allowing domain-specific modules, backends, mechanisms, datasets, and evaluation targets.The architecture is a module map rather than a fixed simulator blueprint.
  • Agents: Agents are stateful economic decision processes that observe permitted signals, form expectations, generate typed actions, and receive feedback from the world.Agent specifications include roles, objectives, internal state, information channels, action spaces, constraints, beliefs, memory, tools, and skills.
  • Runtime protocol: The reusable protocol distinguishes specification interfaces for configurable world components from runtime interfaces for agent actions, world transitions, co-evolution, alignment, and evaluation.The public execution sequence uses ewm.make, world.reset, world.run_agents, world.step, world.coevolve, world.align, and world.evaluate.
  • Economic environments: The environment stores economic state, publishes observations, checks feasibility, orders actions when needed, and applies mechanisms that transform feasible actions into the next state.Mechanisms define participation, admissible action types, price or allocation computation, and settlement of cash and assets.
  • Co-evolution: Unlike static simulations, EWMs allow agents and environments to adapt under explicit controls, representing learning, expectation revision, strategic adjustment, mechanism recalibration, and rule refinement.This co-evolution makes economies evolving systems rather than one-shot input–output mappings.
  • Empirical alignment: Online alignment continuously compares simulated states with real-world signals, detects deviations, and applies bounded corrections to prevent coherent but empirically inconsistent closed simulations.The alignment layer provides the empirical correction channel required for economic digital twins.

5 The Engineering Path Toward Economic World Models

The engineering path toward Economic World Models progresses from designing economic signals and data infrastructure to prompting and contextualizing language-model agents, and ultimately to engineering executable economic environments. Across the six capability levels, later waves increasingly shift from agent-side improvements toward interactive, co-evolving, and empirically aligned economic worlds.

  • Engineering trajectory: Five engineering waves—feature, data, prompt, context, and environment engineering—trace the technical trajectory toward Economic World Models.The waves provide a complementary engineering view to the six-level capability hierarchy.
  • Environment engineering: Environment engineering turns models into worlds by specifying agents, markets, institutions, rules, and state transitions across all six levels.Its role deepens from fixed rules at L1 and feedback channels at L2 to LLM execution interfaces at L3, co-evolution loops at L4–L5, and simulation-to-reality alignment at L6.
  • Feature engineering: Feature engineering forms the observation layer by converting raw economic traces and domain knowledge into structured signals for agents and predictive models.Applications include asset pricing, credit-risk and fraud detection, and signals extracted from text and images.
  • Data engineering: Data engineering shifts the focus from hand-designed features to large-scale data collection, cleaning, alignment, and pretraining infrastructure for stronger predictive and reasoning models.Examples include nonlinear macroeconomic representations, dynamic stock graphs, financial language models, and BloombergGPT’s 50-billion-parameter model.
  • Prompt engineering: Prompt engineering introduces language-based cognitive interfaces through few-shot examples, role-playing, and chain-of-thought prompting, enabling role-aligned economic agents from L3 onward.It remains foundational throughout L3–L6.
  • Context engineering: Context engineering grounds LLM reasoning through retrieval, long-context inputs, memory, reflection, tools, and real-time information, progressing from L3 memory and tool use to L6 empirical grounding.Compared with prompt-only agents, context-engineered agents are more persistent and better grounded.

6 Applications

EWMs organize economic intelligence around three roles: experimentation for human decision makers, reasoning tools for LLM agents, and interactive reinforcement-learning environments for LLM training. Across these roles, they operationalize counterfactual economic reasoning under heterogeneity, institutions, strategic interaction, and long-horizon feedback.

  • Sandbox for human decision making: EWMs provide sandboxes for testing policies, market designs, firm strategies, and research hypotheses before real-world deployment, revealing adaptation, feedback, spillovers, and unintended consequences.They support ex ante evaluation of tax reforms, transfers, industrial policies, regulation, and macroeconomic shocks, including distributional and regime-dependent effects.
  • Tool for LLM agents: For LLM agents, EWMs support counterfactual reasoning and long-horizon planning by simulating candidate actions and endogenous responses within markets, institutions, and organizations.They can compare downstream consequences for trading, procurement, and policy decisions while accounting for prices, incentives, institutional constraints, accounting identities, and delayed feedback.
  • Tool for LLM agents: EWMs provide safety testbeds for evaluating LLM agents before deployment by exposing interaction-driven risks such as collusion, instability, manipulation, harmful equilibria, and cascading failures.These risks can emerge from many agents interacting under realistic institutional and financial constraints rather than from individual errors alone.
  • RL training environment for LLMs: For LLM training, EWMs offer interactive reinforcement-learning environments where agents learn through economic interaction using states, actions, transition dynamics, and short- and long-term rewards.Agents can learn, compete, coordinate, and co-evolve while studying emergent patterns including specialization, inequality, market power, coordination norms, and endogenous segmentation.
  • Shared value: Across all three roles, EWMs make counterfactual economic reasoning operational in environments shaped by heterogeneity, institutions, strategic interaction, and long-horizon feedback.The roles are experimentation for human decision makers, reasoning for LLM agents, and learning for LLM training.

7 Positioning EWM Systems Among Adjacent Paradigms

EWMs are positioned as an implementation layer that preserves economic consistency while enabling richer agents, mechanisms, qualitative information, and data-driven dynamics. Compared with adjacent paradigms, EWMs combine heterogeneous language-mediated reasoning with endogenous market clearing, stock-flow consistency, and macroeconomic validation.

  • Positioning: EWMs complement rather than replace traditional economic modeling, adding an implementation layer for richer agents, mechanisms, and data-driven dynamics across agent, environment, and application layers.The comparison is architectural rather than jurisdictional, and Table 5 summarizes differences across these layers.
  • Traditional economic models: Traditional models achieve consistency through equilibrium and rational expectations, whereas EWMs use explicit heterogeneous agents, observable beliefs, language-based memory, and qualitative observations alongside numerical variables.EWMs retain internal consistency, counterfactual policy analysis, and empirical alignment while exposing parts of the reasoning chain for inspection.
  • Agent-based models: ABMs share heterogeneity, explicit market interactions, stock-flow accounting, and emergent regularities with EWMs, but typically constrain decisions to predefined rules over numerical states.EWMs instead pair structured variables with language-based memory and process quantitative signals and qualitative context in one reasoning step.
  • LLM-based social simulators: LLM-based social simulators target attitudinal and behavioral outcomes, while EWMs make the environment a closed clearing system that endogenously produces prices and allocations.EWM validation moves to macroeconomic time series, with structured beliefs driving willingness prices and actions.
  • World models: EWMs share world models’ ambition for time-evolving learned dynamics and environments usable for downstream-agent training, but differ from single-agent sensory latent-space architectures.Most influential world models center on one embodied agent, high-dimensional sensory streams, compressed latent states, and end-to-end learned dynamics.

8 Discussion and Open Challenges · A LLM-Assisted Literature Collection and Classification

Faithful, evolving, reality-aligned Economic World Models remain a research agenda with open challenges spanning behavioral realism, economic closure, co-evolution, scalability, evaluation, and economic validity. The paper frames these challenges as a CS and AI engineering agenda while leaving equilibrium and counterfactual-validity theory to the EWM/DDGE framework.

  • 8 Discussion and Open Challenges: Faithful, evolving, reality-aligned EWM systems remain a substantial open research challenge rather than a finished technical recipe.The paper presents implementation as a research agenda with several central open challenges.
  • 8 Discussion and Open Challenges: Behavioral realism requires agents to perceive information, form beliefs, respond to incentives, face constraints, adapt over time, and align with real behavioral data.Plausible narratives or rational choices alone are insufficient for resembling real economic actors.
  • 8 Discussion and Open Challenges: Economic closure requires aggregate prices, allocations, risks, liquidity, employment, production, credit flows, and macroeconomic conditions to emerge feasibly from agent interactions.Nonlinear strategic interaction, market clearing, network spillovers, bottlenecks, and feedback loops complicate aggregation.
  • 8 Discussion and Open Challenges: Co-evolution requires both agents and economic rules to change, while controlling what evolves, when it evolves, and whether dynamics remain realistic and stable.Agents may learn and revise strategies, while policies, regulations, contracts, platforms, and institutions respond to aggregate outcomes.
  • 8 Discussion and Open Challenges: Computational and engineering scalability is constrained by heterogeneous interactions, long-horizon rollouts, endogenous mechanisms, calibration, counterfactual validation, and increasingly complex reasoning and communication.Resource costs grow with agents, time steps, agent reasoning, memory, communication, and institutional mechanisms.
  • 8 Discussion and Open Challenges: World-level evaluation must assess both realistic agent beliefs, choices, and adaptation and plausible prices, allocations, distributions, networks, crises, and policy responses.Counterfactual worlds lack a single groundtruth trajectory, requiring evaluation at agent and world levels.
  • 8 Discussion and Open Challenges: Economic validity requires disciplined counterfactual evaluation because behavior can change retraining data and produce the wrong learned environment despite sophisticated agents and empirical alignment.DDGE-style closure supplies an economic fixed-point discipline complementary to the systems architecture.
  • 8 Discussion and Open Challenges: The paper defines a CS and AI engineering agenda for building, evaluating, and governing EWM systems while leaving equilibrium and counterfactual-validity theory to the EWM/DDGE framework.This positions the implementation roadmap as complementary to the framework’s theoretical foundations.

A.1 Data Sources and Search Strategy · A.2 Three-Stage Classification

The review used a systematic, interdisciplinary collection and filtering protocol drawing on arXiv and Web of Science, followed by a three-stage pipeline that classified papers against Economic World Model criteria and six hierarchical levels. The process combined broad keyword retrieval, recall-oriented structural screening, precision-oriented full-text evaluation, and manual checks for borderline cases.

  • A.1 Data Sources and Search Strategy: The literature review implemented a systematic data collection and filtering protocol for an interdisciplinary Economic World Model corpus.
  • A.1 Data Sources and Search Strategy: ArXiv searches covered eight categories spanning economics, finance, artificial intelligence, multi-agent systems, and statistical applications.The categories were econ.GN, econ.TH, econ.EM, cs.MA, cs.AI, cs.LG, stat.AP, and q-fin.ST.
  • A.1 Data Sources and Search Strategy: Web of Science searches complemented preprints with high-impact peer-reviewed research from the UTD-24 journal list.The searches covered premier publications in economics, finance, management, and information systems.
  • A.1 Data Sources and Search Strategy: A three-stage classification pipeline distilled the candidate pool into the final Economic World Model taxonomy.
  • A.1 Data Sources and Search Strategy: Keyword-based pre-filtering used titles and abstracts to assemble a broad repository of papers on economic modeling and agent-based simulation.
  • A.1 Data Sources and Search Strategy: Recall-oriented LLM screening retained papers with heterogeneous agents, interaction-derived endogenous outcomes, and dynamic feedback loops.Generic economic or AI papers lacking integrative modeling were discarded.
  • A.1 Data Sources and Search Strategy: Precision-oriented full-PDF classification evaluated remaining papers against EWM engineering desiderata and assigned valid EWMs to six hierarchical levels.Borderline cases were flagged for manual author verification to improve robustness.

A.2.1 Stage 1: Keyword-based Pre-filtering

Stage 1 used intersection-based keyword matching in paper titles and abstracts, retaining papers containing at least one term from both economic/market and dynamic-modeling or agent-related groups. This process produced a deduplicated pool of 7,836 candidate papers from arXiv and UTD-24.

  • Filtering criterion: Papers were retained when titles or abstracts contained at least one relevant term from each of two thematic keyword groups.The groups covered economic, financial, and market themes alongside dynamic modeling, simulation, multi-agent, reinforcement learning, and world-model themes.
  • Keyword groups: The economic and market keyword group included terms such as “economic,” “financial,” “finance,” “market,” and “macroeconomic.”These terms defined the first side of the intersection-based filter.
  • Keyword groups: The dynamic-modeling and agent keyword group included terms spanning simulation, world models, digital twins, agents, heterogeneous agents, endogenous processes, LLM-based economic agents, and reinforcement learning.It also included dynamic-system, state-transition, and related multi-agent variants.
  • Filtering results: 7,836 candidate papers formed the final deduplicated stage-one pool, comprising 6,008 unique arXiv papers and 1,828 UTD-24 papers.Before deduplication, keyword matching identified 6,028 arXiv candidates and 1,828 UTD-24 papers; the repositories overlapped.

A.2.2 Stage 2: Recall-oriented Screening

Stage 2 used GPT-5.4-mini for recall-oriented screening based only on titles and abstracts, using a low inclusion threshold to minimize false negatives before full-PDF validation. The process retained 794 arXiv papers and 87 UTD-24 papers for later review.

  • Screening strategy: GPT-5.4-mini screened candidate papers using only titles and abstracts, prioritizing recall and deferring rigorous validation and classification to full-PDF review.The low inclusion threshold was intended to minimize false negatives.
  • Retention criteria: The first-stage decision retained papers with either a strong EWM signal or concrete structural evidence of agents generating economic outcomes through dynamic interaction.Structural evidence required economic agents, endogenous outcomes, and dynamic transition, repeated interaction, feedback, or simulation over time.
  • Exclusion criteria: Generic economic topics, forecasting, econometrics, single-agent decisions, static theory, optimization, isolated AI applications, generic simulators, and non-economic ABMs were insufficient evidence by themselves.The screening distinguished EWM-like structures from papers using related terminology without the required economic-agent dynamics.

A.2.3 Stage 3: Precision-oriented Full-PDF Classification

Stage 3 used GPT-5.5 to classify candidate papers from complete PDF evidence with precision-oriented, non-inferential screening. It first validated the EWM definition, then assigned validated papers to one of six levels using agent capability, institutional evolution, and sim-to-real alignment.

  • Full-PDF evaluation: GPT-5.5 evaluated retained candidate papers using complete PDFs and concrete evidence from frameworks, methodologies, simulations, results, and appendices, avoiding unsupported inference.The procedure was explicitly precision-oriented and full-text rather than title- or abstract-based.
  • Sequential classification: The classification used two sequential tasks: binary EWM validation followed by granular assignment to an EWM level.Level assignment was applied only to papers successfully validated as EWMs.
  • EWM binary validation: Four conditions were required simultaneously: an economic resource-allocation world, heterogeneous interacting agents, dynamic transitions, and partly endogenous outcomes generated through agent interactions.Failure to meet any condition excluded the paper.
  • Hierarchical level classification: Validated papers received Levels 1–6 based on agent capability, endogenous institutional evolution, and simulation-to-reality alignment.The agent-capability axis ranges from rule-based to self-evolving LLM substrates.
  • Structured output: The classifier returned structured JSON containing decisions, confidence scores, manual-review flags, exclusion reasons, extracted evidence, and borderline-case explanations.This output was designed to support human validation and prompt calibration.

A.3 Human Validation … B.5 Level 5

The paper validates borderline classifications through full-text human review using the same EWM criteria as the LLM, then organizes representative systems from fixed rule-based worlds through endogenous institutional evolution. Across Levels 1–5, adaptation progressively deepens from fixed rules to agent learning, language-based cognition, persistent agent evolution, and evolving institutions, while real-world alignment remains absent.

  • A.3 Human Validation: Human reviewers read every flagged paper in full and applied the same EWM definitions, exclusion rules, and hierarchical gating logic as the Stage 3 LLM.This protocol was designed to maintain methodological consistency across manual validation.
  • A.3 Human Validation: Reviewers made the final EWM inclusion and level-assignment decisions, extracting missing evidence and overriding tentative LLM classifications when necessary.The process targeted borderline, ambiguous, or low-confidence cases.
  • B.1 Level 1: Level 1 systems generate prices, allocations, and macroeconomic patterns endogenously from decentralized interaction, but specify agent rules, environment dynamics, and governing mechanisms ex ante.These worlds have genuine agent–aggregate feedback but no in-world adaptation, explicit cognitive beliefs, or evolving institutions.
  • B.2 Level 2: Level 2 agents revise strategies from economic outcomes through symbolic mechanisms such as reinforcement learning, evolutionary search, score switching, or Bayesian updating over fixed parameters.Their institutions remain fixed, and agents lack rich internal beliefs, contextual reasoning, and cognitive substrates.
  • B.3 Level 3: Level 3 uses fixed large language models that let agents reason in language, maintain explicit beliefs, recall histories, and reflect on past interactions to adapt in-world.The language model’s weights and skill set remain unchanged across the rollout, despite richer cognition.
  • B.4 Level 4: Level 4 agents acquire persistent strategies, skills, or behavioral routines during simulation, changing the cognitive repertoire that drives later decisions.Implementations include growing strategy libraries, evolving strategic guidance, fine-tuned strategies, and long-term semantic profiles; market rules remain fixed.
  • B.5 Level 5: Level 5 makes institutions endogenous: tax schedules, contracts, governance structures, lending networks, and social norms evolve through agent-driven dynamics.Representative systems study co-evolving policy, contracts, networks, and norms, while the cognitive layer remains comparatively modest.
  • B.5 Level 5: Level 5 still lacks repeated alignment with the real economy because institutions evolve inside the model without updating the model from new empirical observations.This distinguishes endogenous institutional worlds from higher-level sim-to-real economic twins.

B.6 Level 6

Level 6 is the sparsest capability tier: only two works achieve repeated in-run alignment with real economic data, and neither fully realizes a live Sim-to-Real Economic Twin. Both assimilate empirical feedback during model fitting rather than continuously during simulation rollouts.

  • Level 6 scope: Only two works satisfy Level 6’s requirement of repeated, in-run alignment with the real economy.Wiesinger et al. (2010) repeatedly re-optimize heterogeneous minority-game traders against rolling windows of real Nasdaq data, while Evans et al. (2025) formalize adaptive ABM as a Stackelberg game.
  • Level 6 limitations: Neither work fully achieves a Sim-to-Real Economic Twin because empirical feedback enters during model fitting rather than as a live signal during rollout.The cited mechanisms are rolling parameter re-estimation and outer-layer recalibration of latent variables, not continuous absorption of real-world observations that reshapes ongoing agent behavior and beliefs.
Loading 2608.06020v1…