Textbook of AI

Also known as: CoT

Chain-of-Thought (CoT) prompting is a technique that dramatically improves large language models' performance on multi-step reasoning tasks by having them produce intermediate reasoning steps before a final answer. Instead of asking "What is 47 × 29?" and expecting an immediate answer, a CoT prompt asks the model to "think step by step": "First, 47 × 30 = 1410. Then subtract 47: 1410 − 47 = 1363." The explicit reasoning process yields much better accuracy on arithmetic, logic, word problems, and other tasks requiring sequential thought.

Wei et al. (2022) formalised the approach and showed it produces dramatic improvements on arithmetic, commonsense, and symbolic reasoning benchmarks—but only for sufficiently large models. Chain-of-thought is an emergent capability that appears abruptly as models scale beyond certain sizes; smaller models fail to exploit CoT prompting and may even perform worse. This pattern is typical of large language model emergent behaviours.

Extensions include zero-shot CoT (just append "Let's think step by step" without examples), self-consistency (sample multiple chains and take the majority answer), tree-of-thought (explore multiple branches of reasoning), and least-to-most prompting (decompose a problem into subproblems solved in sequence). Chain-of-thought has become a standard tool in LLM deployment and has motivated training techniques that encourage models to produce reasoning chains by default. More recent "reasoning models" like OpenAI's o1 use extended chain-of-thought during training and inference to achieve state-of-the-art performance on scientific and mathematical problems.