En-AgentSociety: Observable and Reproducible LLM-Based Urban Simulation at Regional Scale
Year: 2026
Type: conference-paper
Abstract: LLM-based urban simulators can generate semantically rich cityscale routines, but their outputs are difficult to inspect, reproduce, and validate. We present En-AgentSociety, short for Enhanced AgentSociety, an observable and fault-tolerant simulation platform built on top of AgentSociety. The system records prompts, responses, agent states, memories, emotions, visited places, POI categories, execution times, LLM latency, token usage, and generated trajectories. It also supports scalable regional map generation, traffic simulation integration, and direct export to mobility validation pipelines. In this demo, researchers can run regional simulations, inspect individual agent decisions through a web interface, debug unrealistic mobility behavior, resume failed runs, and compare generated traces against empirical mobility metrics. En-AgentSociety provides a transparent foundation for reproducible LLM-based urban simulation.
Source: HAL
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
Funder: Agence Nationale de la Recherche
OpenAlex ID: W7212364730
