Tree of Thoughts: Deliberate Problem Solving with Large Language Models

Tree of Thoughts (ToT) is a new inference framework for large language models that addresses their reliance on token-level, left-to-right decision making, which falls short on tasks requiring exploration, strategic lookahead, or pivotal early decisions. ToT generalizes the popular Chain-of-Thought prompting approach by enabling exploration over coherent units of text called 'thoughts', which serve as intermediate steps toward solving a problem. The model performs deliberate decision making by considering multiple reasoning paths and self-evaluating its choices, and it can look ahead or backtrack to make global choices.

Experiments on three novel tasks that require non-trivial planning or search—Game of 24, Creative Writing, and Mini Crosswords—show that ToT significantly enhances problem-solving ability. The most striking result is on Game of 24, where GPT-4 with chain-of-thought prompting solved only 4% of tasks while ToT reached a 74% success rate, demonstrating the value of structured search and self-evaluation over linear reasoning; the authors released all prompts and code.