Tree of Thoughts adds search to model reasoning

Shunyu Yao, Karthik Narasimhan and colleagues at Princeton and Google DeepMind generalised chain-of-thought prompting into a search. Where chain-of-thought produced a single line of reasoning left to right, Tree of Thoughts let a model generate several candidate intermediate steps, evaluate how promising each looked, and explore the tree of possibilities — branching forward on good paths and backtracking from dead ends.

The gains were largest on problems that reward planning and exploration rather than a single fluent pass. On Game of 24, an arithmetic puzzle, the authors reported that GPT-4 using chain-of-thought solved 4% of instances, while the same model wrapped in Tree of Thoughts solved 74%. They reported similar improvements on creative writing and crossword tasks. The method used the model itself both to propose steps and to judge them, wiring classical search — breadth-first and depth-first exploration — around the language model as the component that generated and scored moves.

The work was part of a broader shift from treating a model’s answer as a single forward pass toward treating inference as something that could be spent deliberately: sampling more, evaluating intermediate states, and searching. That framing — more computation at inference time bought better answers — anticipated the reasoning models that later built long, self-evaluated deliberation directly into training rather than eliciting it through prompting scaffolds.