LLMs boost urban traffic simulations with adaptive route replanning agents

Urban mobility simulations have traditionally relied on rule-based multi-agent systems that struggle with dynamic environments. Researchers from Brazilian universities propose a novel hybrid architecture integrating Large Language Models (LLMs) as a cognitive decision layer. The system connects the GAMA simulation platform to an external LLM via API, allowing agents to decide whether route replanning is necessary based on current conditions—rather than hardcoded heuristics. A persistent memory module stores past interactions, enabling agents to maintain behavioral consistency over time.

The team compared rule-based and LLM-assisted approaches across multiple road-blockage scenarios and population scales. Results show LLM-enabled agents exhibit significantly greater adaptability and contextual awareness, particularly in environments with higher route flexibility. Memory influenced both performance and behavioral consistency, though effects varied across configurations. The paper positions LLMs as complementary cognitive layers that enrich behavioral representations in spatial multi-agent systems, with implications for more realistic traffic modeling and urban planning.