Guided Domain Solver: Structured Exploration of Domain-Specific Tasks with Large Language Models

ICLR 2026 Conference Submission16862 Authors

Keywords: Monte Carlo Tree Search, Knowledge Graph, Large Language Model, agents, planning

TL;DR: This work utilizes Monte Carlo Tree Search with Knowledge Graphs and Large Language Models to enable sample-efficient problem solving in complex domains.

Abstract: This work presents a method to solve domain-specific problems by leveraging Monte Carlo Tree Search (MCTS), Knowledge Graphs and Large Language Model (LLM) agents. At the core of this approach lies a MCTS algorithm, which explores the complex solution space of a given domain in a goal-directed and sample-efficient manner. In the expansion phase of the MCTS, a domain-specific knowledge graph is incorporated to encode concepts, relationships and constraints. This structured representation enables an LLM agent to make informed decisions for the node expansion. By combining a structured search of the solution space through MCTS, a representation of domain knowledge through the knowledge graph and the generalization abilities of an LLM agent, this method can solve complex tasks in domains where both creativity and adherence to expert rules are essential. In a first step, this approach is used to solve Sokoban, a puzzle game that requires planning and creativity to place several boxes at specific targets with as few moves as possible.

Primary Area: applications to robotics, autonomy, planning

Submission Number: 16862