Tree of thoughts
Era 5 · Generative AI, reasoning, and agents (2022–2026) › Alignment and reasoning
- Lineage
- Symbols and logic
- Kind
- Publication
Who
Princeton University, Google DeepMind
What happened
Shunyu Yao and colleagues propose tree of thoughts: the language model explores several reasoning paths, evaluates its intermediate steps, and can look ahead or backtrack. On the Game of 24, GPT-4 goes from solving 4 percent of the problems to 74 percent.
Why it mattered
It joins language models with classic AI search: reasoning stops being a single line and becomes a tree of alternatives.
The full story
A language model generates text from left to right, word by word, without going back. Even with chain of thought, it follows a single path: if the first step is wrong, everything else is built on top of it. For problems that call for exploring, planning, or correcting, that is a serious limitation.
Yao and colleagues took an old idea from AI, search through a space of possibilities, and applied it to the model's thoughts. Each intermediate reasoning step is a node. The model proposes several possible steps, evaluates for itself which ones look promising, and decides which to pursue, with the option of looking several steps ahead or backing up if a path leads nowhere.
They tested it on three tasks that require planning: the Game of 24, in which four numbers must be combined with arithmetic operations to make 24; creative writing with constraints; and small crosswords. On the Game of 24, GPT-4 with chain of thought solved 4 percent of the problems, and with tree of thoughts, 74 percent.
The method required many more calls to the model per problem than a single chain of thought. But it anticipated the idea that would dominate the following years: giving models more computation at answer time so that they explore and check before deciding.
Three diagrams. A line of four nodes, the chain of thought. A tree that branches into several nodes, in which only one path reaches the bottom, the tree of thoughts. A cycle of three boxes, think, act, and observe, that loops back to think: ReAct.
Builds on
Chain of thought (January 28, 2022)
Documented direct influence
Tree of thoughts "generalizes over the popular Chain of Thought approach" (abstract)
ReAct (October 6, 2022)
Inherited, with no proven direct link
Both papers take the model off a single line of reasoning, one with actions and the other with search
Threads that pass through here
- From prompted reasoning to trained reasoning (step 3 of 9)
Sources
Original source
- Tree of Thoughts: Deliberate Problem Solving with Large Language Models (arXiv:2305.10601) (opens in a new tab). Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Thomas L. Griffiths, Yuan Cao, and Karthik Narasimhan, 2023. open license: CC BY 4.0
Further reading
How to cite
“Tree of thoughts.” The History of AI, revised November 13, 2026. https://thehistoryofai.com/en/milestone/tree-of-thoughts
This entry was last reviewed on November 13, 2026.