Integrating Heuristics and Tool Use into LLM-Based MAPF Workflows

Develop a principled approach to incorporate heuristics and external tool use into large language model-driven multi-agent path finding workflows, specifying how to couple stepwise LLM action recommendations with path-planning or checking modules while preserving per-step coordination constraints and improving solution success rates.

Background

The paper’s current workflow relies on an LLM to propose stepwise actions and a simple rule-based conflict checker to ensure collision avoidance. No heuristics or external planners are integrated, partly because MAPF’s per-step coordination demands make decoupling difficult: a tool that fully accounts for inter-agent conflicts effectively solves the problem without the LLM, whereas tools that ignore coordination cannot ensure valid multi-agent solutions.

The authors explicitly acknowledge uncertainty about how to include heuristics or tool use, noting this as a challenging but meaningful direction. A concrete research need is to design integration strategies that leverage tools to enhance reliability and efficiency without undermining the unique coordination requirements intrinsic to MAPF.