01What chain-of-thought prompting is

Chain-of-thought (CoT) prompting asks a language model to produce a series of intermediate reasoning steps before giving its final answer, rather than jumping straight to the answer (Wei et al., 2022). The idea is simple but powerful: a tricky multi-step problem is far easier to get right if you work through it one piece at a time — and the same turns out to be true for the model. Instead of compressing all the reasoning into a single leap, the model "shows its working," and the act of writing out those steps tends to make the final answer more reliable on problems that need arithmetic, commonsense, or symbolic reasoning.

An important caveat from the original research: this benefit was reported to emerge with model scale — smaller models of that era gained little, and could even be hurt, while sufficiently large models improved markedly (Wei et al., 2022). Treat that as a finding about the models of its time, not a universal law; today's instruction-tuned and reasoning-trained models behave differently.

• CoT = intermediate reasoning steps first, final answer last, instead of answering directly.
• It helps most on multi-step problems — arithmetic, logic, commonsense chains.
• The original "emerges with scale" finding is tied to 2022-era models; don't overgeneralize the threshold (Wei et al., 2022).

02Two ways to ask: few-shot vs. zero-shot

There are two main ways to elicit a chain of thought, and they differ in how much you have to write into the prompt.

A later refinement, Auto-CoT, removes the hand-written exemplars entirely: it clusters questions for diversity and uses the zero-shot trigger to generate reasoning chains automatically, then reuses those as demonstrations (Zhang et al., 2022). Vendor guidance describes a similar ladder — from a plain "think step-by-step," to a guided outline of the steps, to structured tags that separate the reasoning from the final answer (Anthropic docs).

03See it work: direct vs. chain-of-thought vs. self-consistency

Here is one word problem solved three ways. Direct answer jumps straight to a guess — and gets it wrong. Chain-of-thought builds the steps one at a time and lands on the right number. Self-consistency samples several independent reasoning chains and takes the majority vote over their final answers, so one bad chain doesn't decide the outcome (Wang et al., 2022). Pick a mode and run it. The walkthrough below is an illustration of the technique, not output from any specific model.

• Direct compresses everything into one leap — easy to drop a step and answer confidently but wrong.
• Chain-of-thought writes the intermediate steps, so each small calculation can be right before the next.
• Self-consistency samples multiple chains and takes the majority answer, smoothing out one-off mistakes (Wang et al., 2022).

04Why sampling many chains helps: self-consistency & supervision

A single chain of thought can still go wrong — one slip early on and the whole answer is off. Self-consistency addresses this by sampling several diverse reasoning paths for the same prompt and choosing the final answer by majority vote, replacing the usual single greedy path (Wang et al., 2022). Different chains may take different routes, but correct reasoning tends to converge on the same answer, so the majority is more often right than any one chain.

A related thread looks at how we reward reasoning. Process supervision gives feedback on each intermediate step, while outcome supervision rewards only the final answer; rewarding the steps was found to outperform rewarding only the outcome on the MATH dataset (Lightman et al., 2023). And STaR shows models can bootstrap their own reasoning — generate rationales, keep the ones that reach correct answers, and fine-tune on them (Zelikman et al., 2022). These ideas point toward a bigger shift: modern reasoning models (such as OpenAI's o1) produce an internal chain of thought learned via reinforcement learning, rather than relying only on a clever prompt (OpenAI, 2024).

• Self-consistency: sample multiple chains, take the majority-vote answer (Wang et al., 2022).
• Process vs. outcome supervision: rewarding each step beat rewarding only the final answer on MATH (Lightman et al., 2023).
• Prompted CoT ≠ trained reasoning: reasoning models learn an internal chain of thought via RL, a different mechanism from prompt-elicited CoT (OpenAI, 2024).

05Beyond a single chain: the reasoning family

Chain-of-thought is the starting point, not the whole story. Several techniques extend the idea — by decomposing the problem, branching the search, or letting the model act and critique itself. Switch between them to see the core idea of each.

A note on faithfulness. The visible "thinking" or "reasoning" a model shows you is a controlled output or summary — it is not guaranteed to be a complete, literal record of the computation inside the model. Treat exposed reasoning as a useful aid, not proof of exactly how the answer was reached.

06Knowledge check