Exploring several lines of reasoning as a branching structure, evaluating them, and backtracking from the ones that look wrong, rather than committing to a single chain. It suits problems where you can tell a partial solution is going nowhere, such as puzzles and planning.
Technical view
Maintains a search tree over intermediate reasoning states with a generator producing candidate next steps and an evaluator scoring partial states, explored by breadth-first, depth-first or beam search. Substantially more expensive than chain of thought, by an order of magnitude or more, and dependent on a usable state evaluator. Much of its benefit is now obtained internally by reasoning models trained to search during generation.
Related
Part of AI Terminology Index, section: Controlling and Extending Behaviour at Run Time.