From a line to a tree
A chain of thought walks one reasoning path. Tree of thoughts instead branches: at each step the agent generates several candidate next thoughts, scores them, and explores the promising ones, like a search over ideas.
How it runs
- Expand: from the current node, propose a few different next steps.
- Evaluate: score each candidate for how likely it leads to a good answer.
- Select: keep the top branches and prune the rest.
- Repeat down the tree until a branch reaches a solution.
This is a deliberate search, often breadth first or with a value heuristic, rather than committing to the first idea.
Why it wins
- Hard puzzles have dead ends; branching lets the agent back out.
- Scoring partial paths catches bad directions early.
- Diversity at each step raises the chance one path works.
The cost
Every branch is more model calls, so the tree grows expensive fast. Tight pruning and a shallow depth keep it affordable. Tree search helps most on problems with clear intermediate checks, such as math or planning puzzles.
Key idea
Tree of thoughts turns reasoning into a search that branches into several candidate steps, scores them, and prunes weak paths, which solves puzzles a single chain cannot but costs many more model calls.