Tree of Thoughts
A generalisation of chain-of-thought where the LLM explores multiple reasoning branches in parallel, evaluates them, and backtracks — much like a search engine over thoughts.
Yao et al. (Princeton/Google, 2023). Components: a thought generator (proposes next steps), a state evaluator (scores partial states), and a search strategy (BFS, DFS, beam). On Game of 24, GPT-4 went from 4 % (CoT) to 74 % (ToT) success rate.
Expensive — many model calls per problem. Concretely influenced the design of o1/R1, where the "thinking" phase is a learnt internalised version of ToT. Related: Graph of Thoughts, Algorithm of Thoughts.