MVP-MCTS: Language Modelling a Flexible Multi-Agent Planning Framework

Large language models have come under the spotlight in recent years for their seemingly multifaceted capabilities which extend far beyond text processing. In particular, they have been shown to possess logical and reasoning capabilities, which has been augmented in various ways via the use of inference frameworks such as reasoning trees and planning graphs. Meanwhile, some studies tried to explore tasks that are not limited to single-agent scenarios, focusing on tasks in multi-agent settings. However, most of their works focus on text-based role-playing games, while long-term planning games have yet to be extensively explored. In this work we extend the reasoning and planning abilities of the language model to coordinate between multiple agents which have been tasked to achieve specific goals with the least amount of resource expenditure. We devise the framework MVP-MCTS: Multi-agent Value-coordination and Prompt-based Monte-Carlo Tree Search(MCTS) that integrates the prompting of a language model with a tree search procedure without the need for additional grounding. Our method relies on task-specific prompting and the innate world model of the LLM to perform search space simulation and decision-making. Our experiments show that our framework is able to outperform current state-of-the-art LLMdriven planning frameworks.