Suppose you are choosing your next move in a board game. Instead of analyzing every possible continuation, you try each candidate move, play the rest of the game quickly many times, and note how often each move led to a win. Then you focus your remaining time on the moves that look best, while still occasionally checking others in case you misjudged them. That loop of selecting, expanding, simulating and updating the scores is Monte Carlo tree search.
The survey by Browne and colleagues describes MCTS as combining tree search with random sampling, and documents its use well beyond games. DeepMind’s AlphaGo paper in Nature combined MCTS with neural networks that estimated which moves were promising and who was winning, and that system defeated a professional Go player.
Why it matters: MCTS is a clear example of spending extra computation at decision time to get a better answer, an idea that comes up whenever people discuss models that “think longer” before answering. That comparison is editorial; no Chain of Thought episode has discussed MCTS directly.
Sources
- Browne et al., A Survey of Monte Carlo Tree Search Methods (IEEE Transactions on Computational Intelligence and AI in Games, 2012) — Peer-reviewed survey; supports the definition and the selection, expansion, simulation and backpropagation steps, checked October 24, 2026
- Silver et al., Mastering the game of Go with deep neural networks and tree search (Nature, 2016) — Primary paper; supports the AlphaGo description combining MCTS with policy and value networks, checked October 24, 2026