A Six-Level Capability Ladder for 'Economic World Models' Finds Nearly All Existing Work Stuck on the Bottom Rungs

arXiv / HuggingFace Daily Papers (23 upvotes)·low signalXBlueskyLinkedInCopy link

This blueprint paper (arXiv 2608.06020, Aug 6, authors including Dacheng Tao and Lin William Cong) organizes generative economic simulation into six levels running from rule-based agent simulations through LLM-powered adaptive agents to empirically-aligned economic twins, then surveys the literature and finds it concentrated at the low end — systems with self-evolving agents, endogenous institutions and persistent alignment are scarce. The pitch is that such models serve double duty: high-fidelity sandboxes for human decision-makers, and training/planning/evaluation/safety substrates for AI agents. It is a roadmap rather than a result, so its value is as a shared vocabulary for a field that currently lacks one.