CityBehavEx simulates 100K agents for 75 days in under 1 hour on a single GPU
City-scale AI simulations now run on a single consumer GPU—no server farm needed.
City-scale agent-based simulations have long been a trade-off between realism and compute cost. Standard LLM-based approaches can generate rich daily routines but become prohibitively slow when scaled to entire cities. CityBehavEx, developed by Gustavo H. Santos and colleagues, solves this by replacing per-action LLM calls with a hybrid architecture: established human mobility models handle routine movement, while fine-tuned cross-encoders estimate semantic alignment between agent profiles, schedules, and activity transitions. This design cuts inference costs dramatically without sacrificing behavioral depth.
In a validation case study, CityBehavEx simulated 100,000 agents over 75 days in under 60 minutes on a single consumer-grade GPU—a task that would typically require days of server-class compute. The platform also exposes agent trajectories and activity traces for inspection, lets users define simulation regions, and validates outputs against real-world mobility, time-use, and semantic metrics. The result is a tool that researchers and urban planners can use to run large-scale, empirically grounded simulations on their own hardware, debugging unrealistic behaviors in real-time.
- Hybrid approach combines classic mobility models with fine-tuned cross-encoders to avoid expensive per-agent LLM calls.
- Simulated 100,000 agents over 75 days in under 1 hour on a single consumer GPU.
- Platform supports empirical validation against real-world spatial, temporal, and semantic distributions.
Why It Matters
Democratizes large-scale urban simulation—now runs on a desktop GPU with validated realism for planners and researchers.