Tree of Thoughts: Deliberate Problem Solving with Large Language Models

Authors: Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Thomas L. Griffiths, Yuan Cao, Karthik Narasimhan

Published May 17, 2023

Read the paper: arXiv 2305.10601 · PDF

Why it matters

ToT is the reference point for any "search-over-reasoning" technique. It made clear that the bottleneck on hard reasoning wasn't model capability but the greedy linear reasoning chain. Modern reasoning models internalize this -- but if you're implementing an agent, ToT-style exploration remains a useful tool for branching search problems.

How to use it

Use ToT when the task has genuine branching structure (planning, game-like tasks, constrained optimization). For everyday reasoning, self-consistency (sample N, majority vote) is cheaper and nearly as good.

Abstract

We introduce Tree of Thoughts, a new framework that generalizes over chain-of-thought prompting and enables exploration over coherent units of text that serve as intermediate steps toward problem solving.