Cognitive Agents in Urban Mobility: Integrating LLM Reasoning into Multi-Agent Simulations.

Area of Science:

• Urban planning and transportation science
• Artificial intelligence and agent-based modeling
• Cognitive science and behavioral economics

Background:

• Urban mobility systems face sustainability, equity, and resilience challenges.
• Traditional agent-based models (ABMs) lack cognitive depth for simulating adaptive user behaviors.
• Environmental pressures exacerbate urban mobility system complexities.

Purpose of the Study:

• To propose a novel cognitive agent architecture using Large Language Models (LLMs).
• To enhance the realism, flexibility, and interpretability of urban mobility simulations.
• To model adaptive user behaviors in response to transport disruptions.

Main Methods:

• Developed a cognitive agent architecture based on LLMs with memory-driven planning, reflection, and adaptation.
• Integrated LLM agents into the SimFleet agent-based simulator.
• Conducted a 20-day simulation with over 320 individuals and performed an ablation study.

Main Results:

• Observed emergent adaptation patterns in agents under stable and disrupted transport conditions.
• Quantified the impact of short-term and long-term memory modules on agent reasoning.
• Demonstrated LLM-driven agents' ability to dynamically generate, adjust, and reflect on travel plans.

Conclusions:

• LLM-driven cognitive agents significantly enhance the realism and flexibility of urban mobility simulations.
• The proposed architecture offers improved interpretability for user responses to disruptions.
• This approach advances agent-based modeling for complex urban systems.