This paper investigates the challenges of using large language models (LLMs) for solving multi-agent path finding (MAPF), a problem involving coordinating multiple agents to navigate a grid without collisions. While LLMs have shown promise in various tasks, they struggle with MAPF due to the complexity of coordination and planning. The study evaluates LLMs on three standard MAPF benchmark maps: empty, room, and maze. LLMs succeed in simple scenarios but fail in more complex ones, often due to limitations in understanding the environment, context length, and reasoning capabilities. The paper highlights three main issues: 1) LLMs lack the ability to understand and remember obstacle locations, 2) context length limits hinder their performance, and 3) reasoning capabilities are insufficient for complex planning. Experiments show that LLMs can generate valid solutions for small-scale problems but fail when the number of agents increases. The study also explores different input methods, including text-based and image-based prompts, but finds that LLMs struggle with complex environments like mazes. The paper discusses how researchers from different backgrounds can contribute to solving MAPF using LLMs, emphasizing the need for improved reasoning, context handling, and obstacle understanding. It also notes the challenges of using LLMs in real-world scenarios, including latency and the need for efficient planning algorithms. The study concludes that while LLMs have potential, their current limitations in reasoning and context handling make them unsuitable for complex MAPF tasks. Future research should focus on improving these capabilities to enable more effective use of LLMs in multi-agent systems.This paper investigates the challenges of using large language models (LLMs) for solving multi-agent path finding (MAPF), a problem involving coordinating multiple agents to navigate a grid without collisions. While LLMs have shown promise in various tasks, they struggle with MAPF due to the complexity of coordination and planning. The study evaluates LLMs on three standard MAPF benchmark maps: empty, room, and maze. LLMs succeed in simple scenarios but fail in more complex ones, often due to limitations in understanding the environment, context length, and reasoning capabilities. The paper highlights three main issues: 1) LLMs lack the ability to understand and remember obstacle locations, 2) context length limits hinder their performance, and 3) reasoning capabilities are insufficient for complex planning. Experiments show that LLMs can generate valid solutions for small-scale problems but fail when the number of agents increases. The study also explores different input methods, including text-based and image-based prompts, but finds that LLMs struggle with complex environments like mazes. The paper discusses how researchers from different backgrounds can contribute to solving MAPF using LLMs, emphasizing the need for improved reasoning, context handling, and obstacle understanding. It also notes the challenges of using LLMs in real-world scenarios, including latency and the need for efficient planning algorithms. The study concludes that while LLMs have potential, their current limitations in reasoning and context handling make them unsuitable for complex MAPF tasks. Future research should focus on improving these capabilities to enable more effective use of LLMs in multi-agent systems.