chain-of-thought

Chain of Thought (CoT; Wei et al., 2022) — prompting the model to “think step by step” so it spends more test-time computation decomposing a hard task into smaller, simpler steps before answering. Per Weng 2023, CoT became “a standard prompting technique for enhancing model performance on complex tasks”, and it does double duty: it improves accuracy and sheds light on the model’s reasoning process.

The mechanism

• Instead of mapping prompt → answer directly, the model emits intermediate reasoning steps, then the answer.
• This is test-time compute: more reasoning tokens for harder problems, traded for accuracy. (The idea later gets trained in rather than prompted — see reasoning-models.)
• In agent terms, CoT is one way to do task-decomposition: “transform big tasks into multiple manageable tasks.”

Tree of Thoughts (the branching extension)

Tree of Thoughts (ToT; Yao et al. 2023) generalizes CoT from a single chain to a tree:

• Decompose the problem into thought steps; generate multiple candidate thoughts per step.
• Search the resulting tree with BFS or DFS.
• Each state is evaluated by a classifier (via a prompt) or by majority vote.

Where CoT commits to one reasoning path, ToT explores several and prunes — useful when a single greedy chain is likely to go wrong.

Relationship to neighboring concepts

• vs. ReAct — CoT reasoning produces an answer within a turn; ReAct interleaves reasoning with environment-changing actions across turns. ReAct ⊃ CoT in spirit: the “thought” step is CoT-flavored, but it sits in an action loop. (The react page draws this distinction explicitly.)
• vs. reasoning-models — o1 / R1-style models are what happens when CoT stops being a prompt trick and becomes an RL-trained policy over how much to think. Pre-reasoning-model era: CoT was a prompting convention. Reasoning-model era: the model has internalized when to spend reasoning tokens.
• as task-decomposition — CoT, ToT, and "Steps for XYZ.\n1." prompting are all listed by Weng as decomposition mechanisms.

Cross-link

• 2026-06-04-llm-powered-autonomous-agents — the source; CoT/ToT under “Task Decomposition”.
• task-decomposition — the planning strategy CoT serves.
• reasoning-models — CoT trained into the weights as inference-budget allocation.
• react — reasoning-as-action; the loop primitive that subsumes CoT’s “thought”.