Term
Tree of Thoughts
Tree of Thoughts (ToT) is a prompting technique in which an LLM explores multiple solution paths as a branching tree, scores them and follows only the promising branches — a generalisation of Chain-of-Thought.
Tree of Thoughts — explained in more detail
Chain-of-Thought generates one linear reasoning path. Tree of Thoughts goes further: at each step the model produces multiple candidate next steps (“thoughts”), evaluates them itself (e.g. “promising”, “dead-end”) and searches the resulting solution tree with classic strategies like breadth-first or depth-first. Dead-ends are pruned, promising branches are pursued. The technique was introduced by Yao et al. in 2023 and targets tasks where a single attempt often fails and backtracking is needed.
Example / Practical context
On the “Game of 24” puzzle (combine four numbers using basic arithmetic to hit 24) GPT-4 with Chain-of-Thought reaches about 4% success — with Tree of Thoughts the original paper reports around 74%. The cost is steep: ToT can use ten to a hundred times more LLM calls per task, because every node in the tree triggers its own evaluation and expansion calls. In day-to-day use the technique only pays off on hard reasoning or planning problems, not on simple questions.
Distinction from related concepts
Chain-of-Thought is the linear special case without branching. Self-Consistency generates multiple independent CoT paths and picks the most frequent answer — no tree, no per-step scoring. Graph of Thoughts extends ToT with cross-links between nodes and is the next generalisation step.
Entdecke mehr
Effort level and deep thinking: two independent axes for AI tasks
Effort scales breadth, deep thinking scales depth. When each setting makes sense — with three clear examples and one rule of thumb.
GlossarChain-of-Thought
Chain-of-Thought (CoT) is a prompting technique that asks the model to spell out its reasoning in intermediate steps — boosting accuracy on multi-step tasks.
LexikonPrompt Security — Injection, Leaking, Guardrails
How prompt injection, prompt leaking, and jailbreaks work — and which defenses (guardrails, spotlighting, sanitization) actually help.