GenWorld simulates 196K AI agents in realistic Japanese city

LLM-agent simulations have historically struggled with scalability and grounding: agents need environments that reflect real urban constraints, but running city-scale populations with direct LLM calls is computationally prohibitive. GenWorld, a new infrastructure from a team of researchers, solves both problems. It combines a building-level synthetic city (instantiated for Higashihiroshima, Japan) with a structured agent-environment interface and, crucially, an offline compilation step that transforms LLM-derived decision signals into lightweight lookup policies. This enables rollout of 196,608 synthetic residents whose demographics are validated against official census tabulations, and whose commuting patterns are diagnosed using YJMob100K mobile-phone data.

GenWorld demonstrates its capability through three reproducible case studies: a full-city weekday rollout capturing routine behaviors, a weekday-weekend behavioral contrast, and a warning-response perturbation that provides auditable replanning traces. The infrastructure is open-source with code and project page available. While calibrated forecasting for traffic, evacuation, or policy outcomes remains future work, GenWorld already provides a robust platform for grounded, large-scale LLM-agent studies. It lowers the barrier for researchers to experiment with urban AI simulations, enabling realistic agent behavior studies without the compute cost of online LLM calls for every agent decision.