Mon 07:00
0 agents on the road
Legend
on the way to an activity (cabin colour = destination)
at a place
COLM 2026 venue (Hilton Union Square)
A synthetic San Francisco built from OpenStreetMap buildings and Census data, inhabited by 1,000 agents. Every agent is an open-weight LLM (GPT-OSS-120B) that plans its day, decides what to do next each time a stay ends, remembers past days, and reflects every night.
Storm week. On Tuesday and Saturday one sentence is added to every agent's prompt: a storm warning. Nothing else changes. Switch to No storm to see the same agents' week without it.
Track an agent (or click any car or dot) to follow it and read its plans, decisions, memories and reflections, word for word from the simulation logs.
UrbanLLMind: Scalable LLM-Powered Urban Mobility Simulation with Open-Weight Models. Prabin Bhandari, Sandro M. Reia, Dieter Pfoser, Antonios Anastasopoulos (George Mason University), COLM 2026. Contact: pbhanda2@gmu.edu. Map data © OpenStreetMap contributors; shoreline: U.S. Census Bureau TIGER/Line.