Where do economic world models currently stand in capability?

The paper defines Economic World Models (EWMs, after Cong, 2025) as generative models that "simulate how economies evolve from within," modeling heterogeneous agents, their beliefs and actions, and the market and institutional mechanisms through which their interactions produce aggregate outcomes. It then organizes EWM systems into a six-level capability ladder: fixed rule-based agent worlds, adaptive and LLM-based agent worlds, self-evolving agents, evolving institutional worlds, and sim-to-real economic twins aligned with real observations. Its survey finding is 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."

The motivation is a claim about explanation. The introduction contrasts the "observe and explain (or predict)" tradition of economics with a generative one: deriving an equilibrium, fitting parameters, or forecasting a turning point does not explain a phenomenon "unless the outcome can emerge from the model inside." The paper's premise is therefore that understanding an economy means building "a world simulator which generates such outcomes." The ladder ranks systems by how much of the economy is produced from inside the simulation. Agents move from scripted to adaptive to self-evolving, institutions move from given to endogenous, and the whole is finally checked against real observations. The discussion adds the coupling that makes this a world model rather than an agent benchmark: agents act on incentives, information, beliefs and constraints, and their actions are aggregated through economic mechanisms into prices, allocations, risks, institutions and future states.

Against the nearest notes, this is a domain-specific instance of the simulator framing in What should a world model actually be designed to do?, with the difference that the simulated space is a population of interacting agents plus the rules that aggregate them. It organizes the field along a different axis than What five design choices compose a world model?, which splits the design problem by component; the ladder splits it by how endogenous the world is. It also differs from Can language models learn to simulate agent environments?, where a language model predicts environment states. Here the world's dynamics come out of many agents' decisions. The named open challenge of behavioral realism, that agents should not "merely produce plausible narratives or rational choices," overlaps with the steerability concerns in Can we make LLM social simulations interpretable and steerable?.

The excerpt is silent on several points. It lists the levels' contents without saying how the six levels map onto the named stages. It gives no criteria for assigning a system to a level, no count of surveyed works, and no example of a system at the top levels. It reports no validation results. The discussion breaks off after the first open challenge, behavioral realism, so the remaining challenges are not visible here. The paper itself calls EWM implementation "a research agenda" and not "a finished technical recipe," and that framing is the strongest claim the excerpt supports. The ladder is a proposed organizing scheme, and the "lower levels crowded, upper levels rare" finding is the survey's reading of the field, not a measured result.