Tree of Thoughts

Tree of Thoughts is a 2023 research method that explicitly branches into candidate reasoning paths, evaluates them, and backtracks from dead ends.

Tree of Thoughts (ToT) organizes problem solving as a search tree. A controller generates candidate next steps, scores them, continues promising branches, and can return to an earlier branch after a dead end. In a room-scheduling task, one branch might place the first meeting in Room A while another tries Room B; the controller rejects a branch that later violates a constraint. ToT is an explicit search procedure. A model exploring alternatives internally does not establish that an application ran the published method.

Builder example

Each branch consumes model calls and evaluation work. Explicit ToT can help when early choices constrain later choices and the application can score partial progress. Summarization or extraction rarely supplies that branching structure. Deterministic schedulers or search algorithms may also fit constraint problems with exact rules, so compare them before choosing a model-driven tree.

Common confusion: A collection of independent answers is best-of-N sampling. ToT requires intermediate branches, evaluation, and backtracking. A model's private exploration also differs from an application that implements the published tree procedure.