CitySim: Using LLM Agents to Simulate and Predict Urban Populations

Japanese researchers created a system that simulates an entire city by generating up to 1 million virtual residents who behave like humans using LLMs.And it predicted real-world with terrifying accuracy.They created an urban simulator called “CitySim” that populates a digital twin of Tokyo with up to 1 million autonomous AI agents.Each resident is powered by an LLM, complete with:• Personal memories• Long-term goals• Human-like desires (hunger, fatigue, social connection)• Spatial awareness and navigationMost urban simulations rely on rigid, hand-crafted rules, if A happens, B follows.CitySim is different. It’s “value-driven.” The agents have agency. They generate their own schedules based on their personal habits, situational needs, and changing environments.When researchers ran the simulation in Tokyo, the output was chillingly accurate.Weekday commuting patterns, weekend leisure activities, and shopping habits matched Japanese government statistical data with near-perfect precision.It even successfully predicted crowd distributions in Shibuya and identified which stores would actually become popular, months before the real-world data caught up.The implications for urban planning are massive.We can now safely and cost-effectively test disaster preparedness, commercial site planning, and public infrastructure upgrades without moving a single physical person.But it’s not just about planning.The researchers are already looking at implementing high-level human drives, like “self-actualization,” into these agents to see how they evolve.The paper went up on arXiv in June 2025 and almost nobody outside computational social science has read it. Get full access to Factions Of Freedom's Substack at freedomfaction.substack.com/subscribeBecome a supporter of this podcast: https://www.spreaker.com/podcast/factions-of-freedom--6512929/support.