Assessing LLM Capability for Multi-Agent Path Finding

Determine how effective large language models are at directly solving multi-agent path finding on four-connected grid maps without auxiliary solvers or fine-tuning, specifically whether they can consistently generate valid collision-free multi-agent plans and achieve good makespan performance across standard benchmark maps and agent counts.

Background

Multi-agent path finding (MAPF) requires coordinating multiple agents on a four-connected grid to reach individual goals without collisions, typically optimizing makespan. The paper highlights that MAPF’s coordination is tightly coupled with stepwise path planning, making it difficult to decouple high-level guidance from low-level planning in a way amenable to tool use by LLMs. Because of this coupling, whether LLMs can autonomously perform MAPF planning effectively is uncertain.

The authors investigate direct use of LLMs to select actions step-by-step, accompanied only by a high-level conflict checker. They demonstrate promising results on easy, obstacle-free maps, but consistent failure on more complex maps (rooms, mazes), motivating the explicit open question regarding LLMs’ capability to solve MAPF reliably and efficiently.