MCTS-Reasoning: A Canonical Specification of Monte Carlo Tree Search for LLM Reasoning

This technical report provides a rigorous specification of Monte Carlo Tree Search (MCTS) applied to Large Language Model (LLM) reasoning. We present formal definitions of the search tree structure, the four phases of MCTS (Selection, Expansion, Simulation, Backpropagation), and their adaptation for step-by-step reasoning with language models. Key design choices—including tree-building rollouts and terminal-only evaluation—are explicitly documented and justified. The goal is to establish a canonical reference specification that authentically captures the MCTS algorithm while adapting it for the unique characteristics of LLM-based reasoning tasks.