Tree of thoughts
A reasoning pattern where the LLM explores multiple solution paths in parallel as a tree, evaluates partial paths, and backtracks — outperforming linear chain-of-thought on hard problems.
Tree of thoughts (ToT) generalizes chain-of-thought. Where CoT lays out one reasoning trace, ToT branches: at each step, the model generates several candidate next steps, evaluates them with a heuristic or stronger model, and continues down the most promising branches.
The Yao et al. (2023) paper showed ToT solving Game of 24 puzzles at 74% accuracy where CoT scored 4%. The technique generalizes: combinatorial planning, code-generation with verification, math word problems all benefit when the right answer is hard to reach in one pass.
In 2026 production stacks, ToT shows up two ways. Implicitly: reasoning models like o3 and Claude with extended thinking do branchy reasoning internally. Explicitly: agents like Devin and Manus run a planner that generates candidate plans and selects the best one before executing.
Frequently asked
When should I use tree of thoughts over chain of thought?+
ToT wins on problems where the right path is hard to find in one shot — combinatorial planning, complex math, code with many viable paths. For routine tasks, CoT is cheaper and equally effective.
Do reasoning models do tree of thoughts internally?+
Effectively yes — they explore alternative reasoning paths in their internal chain of thought. The mechanism is not explicitly tree-structured, but the practical effect (consider alternatives, evaluate, pick the best) is similar.