Tree of Thought
Tree of Thought (ToT) is a prompting framework for large language models that explores multiple candidate reasoning paths simultaneously, enabling deliberate backtracking and evaluation of intermediate steps rather than committing to a single linear chain of thought.
Tree of Thought is a problem-solving framework introduced by Yao et al. in a 2023 paper from Princeton and Google that extends chain-of-thought prompting. Rather than generating one sequential reasoning trace, ToT structures the model's deliberation as a tree: each node represents a partial solution or intermediate thought, and the model generates several candidate continuations at each node, then evaluates and selects which branches to explore further. This allows the model to consider, compare, and abandon reasoning paths much as a human might sketch and discard partial solutions.
The framework applies classical tree-search algorithms—breadth-first search (BFS) or depth-first search (DFS)—over a space of reasoning steps. At each expansion, the model generates multiple candidate thoughts, then a separate evaluation prompt asks the model to score each intermediate state as promising, uncertain, or unpromising, enabling informed pruning. The process terminates when a satisfactory final answer is found or the search budget is exhausted.
ToT handles problems that require exploration and planning—combinatorial puzzles, multi-step mathematical proofs, structured creative writing—where greedy single-path reasoning frequently fails. The original paper demonstrated substantial gains on the Game of 24 (a mathematical reasoning task) and crossword puzzle completion compared to standard chain-of-thought. The key insight is that evaluation and search, not just generation, are integral to deliberate problem solving.
By 2026, ToT's influence is most visible in reasoning models such as OpenAI's o3 and Anthropic's Claude with extended thinking, which perform analogous internal exploration natively during inference rather than requiring external orchestration. Academic extensions include graph-of-thought and forest-of-thought variants. Explicit ToT orchestration via prompting remains useful in constrained or interpretability-focused settings, but the paradigm has largely been absorbed into end-to-end trained reasoning systems.