Flashcard: what is Tree of Thoughts (ToT) prompting and what problem does it solve?

**Tree of Thoughts (ToT)** is a prompting strategy that lets a language model explore several reasoning paths at the same time instead of committing to one straight chain. With plain chain-of-thought, you ask the model to think step by step and it produces one long reasoning chain followed by an answer. If the first step is wrong, the whole chain is poisoned and the answer is usually wrong too. There is no way for the model to back up. With Tree of Thoughts, at each reasoning step the system generates multiple possible next thoughts (branches), scores each branch (often by asking the same model to rate how promising it is), keeps the best one or two, and discards the rest. From those survivors it generates more branches at the next step. The result is a tree of partial reasoning paths, with the algorithm always continuing from the most promising leaf. The technique is designed for problems where the first move you pick can turn out to be wrong: puzzles, planning, search, and constraint problems. It is overkill for simple Q&A.

**Tree of Thoughts (ToT)** is a prompting framework that turns reasoning from a single chain into an explicit search tree. At each step the system prompts the LLM to generate several candidate next thoughts, scores them with an evaluator (typically the same LLM asked to rate each branch on a scale), prunes the weak ones, and continues from the strongest. The reasoning structure is a tree; the algorithm wrapping it is BFS or DFS with pruning. The motivation is the failure mode of chain-of-thought. CoT commits to one reasoning path and cannot recover if an early step is wrong; the whole chain is contaminated and the final answer is usually wrong. ToT addresses this by exploring multiple paths and using the evaluator to abandon dead ends. The technique shines on problems with **backtracking**: Game of 24 (the original paper's headline benchmark), crosswords, planning, puzzles, constraint satisfaction. On these tasks ToT can substantially outperform CoT. For routine tasks (summarize, classify, translate, answer a factual question) ToT is overkill and burns 10-50x the LLM calls of CoT for no quality gain. The practical knobs are branching factor (how many candidates per step), depth (how many steps deep to search), and evaluator design (how to score branches reliably). ToT is a heavy hammer; reach for it when the task structure actually needs search.