Chain-of-Thought

Why it works

For problems needing multi-step reasoning, demanding the answer directly forces the model to "jump to the conclusion." Having it write the process spreads the computation across many tokens, each step conditioned on the last, lowering the chance that a mid-way error is unrecoverable. Kojima et al. also found that just adding "Let's think step by step" (zero-shot CoT) triggers a similar effect.

Boundary with ReAct

Chain-of-thought is purely internal reasoning: no lookups, no tool calls — just logic unfolded in text, suited to maths, commonsense inference, logic puzzles. ReAct inserts real external actions between reasoning steps. Use ReAct when you need facts or environment interaction; use chain-of-thought when you just need to "think it through."

Common misconceptions

• The written "reasoning steps" are an output format and do not necessarily reflect the model's true internal computation.
• Chain-of-thought helps little on simple factual questions and adds token cost.

Related terms

Sources

1. Wei et al. — Chain-of-Thought Prompting Elicits Reasoning in Large Language Models, NeurIPS 2022 (arXiv:2201.11903)
2. Kojima et al. — Large Language Models are Zero-Shot Reasoners (arXiv:2205.11916)