Researchers present CityBehavEx, an interactive urban simulation platform that combines human mobility models with fine-tuned cross-encoders to generate semantically rich city routines at scale. This approach avoids invoking large language models for every agent action, enabling simulations of 100,000 agents over 75 days in under one hour on a single consumer GPU.
- Integrates established human mobility models with fine-tuned cross-encoders to estimate semantic alignment between agent profiles and activity transitions.
- Generates mobility patterns that better match real-world spatial, temporal, and semantic distributions.
- Allows users to define simulation regions, inspect trajectories, debug behaviors, and validate routines against real-world metrics.
The platform addresses the high costs and weak validation of previous LLM-based simulators by exposing agent behavior for inspection and supporting empirical validation.