A paper turned single-path reasoning into a search over thoughts

The conversation

On 17 May 2023, Shunyu Yao of Princeton and coauthors from Princeton and Google DeepMind posted a paper titled 'Tree of Thoughts: Deliberate Problem Solving with Large Language Models.' Chain-of-thought prompting, already on file, asks a model to write one linear sequence of reasoning steps. This paper's proposal, Tree of Thoughts (ToT), instead treats each intermediate 'thought' as a node a model can branch from, evaluate, and abandon, letting GPT-4 explore several reasoning paths and backtrack when one stalls, rather than commit to its first line of reasoning.

What the documents show

The paper reports results on three tasks using GPT-4. On Game of 24, an arithmetic puzzle, standard input-output prompting solved 7.3% of tasks, chain-of-thought prompting 4.0%, and self-consistency chain-of-thought 9.0%; ToT with a search breadth of one reached 45%, and a breadth of five reached 74%. The authors also tested an oracle best-of-100 chain-of-thought baseline, which reached 49%, still below ToT's wider search. On 5x5 mini crosswords, the paper reports a word-level success rate under 16% for input-output prompting against 60% for ToT, which fully solved 4 of 20 games. For creative writing, GPT-4 itself was used to score outputs for coherence, with ToT rated 7.56 out of 10 on average, a model-judged score the authors report rather than a task with a single correct answer.

The system boundary

ToT is a prompting and search procedure layered on an unmodified language model; it does not retrain or fine-tune GPT-4. The model itself proposes candidate thoughts and also scores them, so evaluation quality is bounded by the same model's judgment, not an independent checker. The paper frames ToT as useful specifically where a task needs exploration, lookahead, or backtracking; for tasks GPT-4 can already solve directly, the authors say deliberate search 'might not be necessary.' Where the model's own thought-scoring is unreliable, as the crossword error analysis shows, the search structure cannot compensate for a bad evaluator.

Where it fails

The authors' own limitations section states this version explores only three relatively simple tasks and that ToT needs substantially more GPT-4 API calls than single-pass sampling, a tradeoff left to the user to manage. Their crossword error analysis separately notes the model's evaluator sometimes prunes a correct word as 'impossible' because GPT-4 fails to recognize an obsolete but valid term, a documented case of the scoring step discarding a correct branch. The project repository, MIT licensed and tied to a NeurIPS 2023 paper, ships the exact prompts used but does not claim the method transfers beyond the three tasks tested.

• Does the task actually need exploration and backtracking, or would a single reasoning pass already solve it?
• Who or what is scoring each intermediate step, and could that same model be systematically wrong about it?
• What is the added inference cost of a wider search, and is it justified by the accuracy gain reported for this task?

Tree of Thoughts is a documented case where widening a search past one reasoning path measurably helped GPT-4 on three specific puzzle-like tasks; it is not evidence the same search strategy pays off wherever an agent needs to plan.