Jiale Han, Xiang Li, Jing Qian, Wenyuan Gu, Pin Gao, Ye Luo, Hongyuan Zha, Dacheng Tao, Benyou Wang, Lin William Cong
6 min
Economic World Models (EWMs) represent a shift from traditional economic modeling—which often relies on aggregate observations and external equilibrium assumptions—to a generative, agent-based approach. The authors argue that to truly understand economic phenomena, we must build simulators where economies evolve from the inside. In these systems, heterogeneous agents (households, firms, banks, regulators) observe information, form beliefs, and interact through market mechanisms, with aggregate outcomes emerging endogenously from these micro-level actions.
The paper introduces a six-level capability ladder to categorize the maturity of EWM systems. This progression moves from basic, rule-based agent environments to sophisticated systems featuring LLM-based agents, self-evolving institutions, and eventually, 'sim-to-real' digital twins that align with empirical data. By treating the EWM as a modular software runtime, the authors provide a concrete execution interface (e.g., world.step(actions)) that allows researchers to treat economic environments as programmable, interactive sandboxes.
As AI agents become increasingly capable of performing complex economic tasks—such as trading, negotiation, and strategic planning—the need for safe, high-fidelity testing environments becomes critical. EWMs offer a substrate for:
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.
Sam: So how do you measure how good a simulator actually is? There must be some way to rank them.
Alex: The authors propose what they call a "six-level capability ladder." At the bottom, you have simple rule-based systems — agents that follow fixed instructions with no real reasoning. As you climb the ladder, agents become more sophisticated: they start learning from experience, then adapting to new situations, and eventually the goal is what the authors call "sim-to-real" digital twins — simulators so well-calibrated to real-world data that their outputs are genuinely useful for policy decisions.
Sam: And I'm guessing we're not near the top of that ladder yet.
Alex: Most current work sits at the lower levels. Systems where agents can evolve their own strategies and stay continuously aligned with real-world data are still quite rare. It's an active frontier, not a solved problem.
Sam: So what's changed that makes this feel achievable now, rather than just a distant ambition?
Alex: Two developments have converged. First, large language models — the kind of AI behind modern chatbots — have become capable of simulating something that looks a lot like human reasoning: weighing options, responding to context, adjusting based on new information. That gives researchers a much more plausible building block for individual agents. Second, the sheer volume of economic data available today is far greater than it was even a decade ago, which means there's more to calibrate and test these models against.
Sam: But that raises an important question. How do you know your simulated economy hasn't quietly drifted away from reality? A model that feels accurate but isn't could be worse than no model at all.
Alex: The authors address this directly. They propose a "real-world alignment" layer — essentially a continuous checking mechanism that compares what the simulation produces against actual empirical data, and corrects for any drift. It's a bit like a ship's navigator constantly checking their position against GPS rather than trusting the compass alone.
Sam: So the ambition here isn't just to build something clever — it's to build something reliable. Something you could genuinely trust when the stakes are high.
Alex: That's the paper's central argument. The authors suggest that the next major step for both economics and AI isn't a more sophisticated equation — it's building the infrastructure to make these simulated worlds modular, testable, and honest about their own limitations. The goal is a tool that policymakers can actually use, not just an academic exercise. And given how costly real-world policy mistakes can be, that's a goal worth taking seriously.
Sam: It makes you think differently about how we make big economic decisions. Right now, we're essentially running experiments on real people. A reliable simulator would change that calculus considerably.
Alex: It would. And that's precisely why the authors frame this as a long-term infrastructure project — not a single breakthrough, but a sustained effort to build the foundations. Thanks for listening to ResearchPod.