Chain-of-Thought Prompting Explained: The Simple Trick Behind Every Reasoning Model

Chain-of-thought (CoT) prompting means getting a language model to write out intermediate steps before it answers. In the original paper (Wei et al., Google Research, Brain Team, arXiv 2201.11903, first submitted 28 January 2022), eight worked examples took PaLM 540B on the GSM8K math benchmark from 17.9% to 56.9%, according to the authors. That edged past the previous best of 55%, a fine-tuned GPT-3 with a verifier.

This post decodes that paper and its immediate follow-ups, then covers where the idea is weakest. All results below are the authors' claims unless a later study is cited. Reading basis: the Wei et al. and Kojima et al. papers were read in their arXiv HTML versions; the other papers were read at abstract level, and numbers taken from them are limited to what their abstracts state.

What problem does chain-of-thought prompting solve?

It targets multi-step problems that large models fail when asked for the answer directly. Standard few-shot prompting shows a model question and answer pairs and asks it to continue the pattern. For arithmetic word problems, that gives the model one forward pass to produce a number.

The Wei et al. authors note that earlier work trained models on written rationales, but that it is costly to create a large set of high quality rationales. Their alternative needs no training. You write a handful of exemplars where the answer is preceded by reasoning, and the model imitates the format.

How does chain-of-thought prompting work?

You put worked reasoning inside the prompt, and the model produces similar reasoning for the new question. The authors manually composed eight few-shot exemplars with chains of thought and reused them across most benchmarks. For AQuA they used four exemplars from the training set.

An exemplar looks like this, in abbreviated form:

Q: A cafe had 23 apples. It used 20 for lunch and bought 6 more. How many now?
A: The cafe started with 23. It used 20, leaving 3. It bought 6 more, so 3 + 6 = 9. The answer is 9.

The models tested included GPT-3, LaMDA, PaLM, UL2 20B and Codex. No weights change. The only difference from standard prompting is the content of the examples.

Can you skip the examples with "Let's think step by step"?

Largely yes, for math. Kojima et al. (University of Tokyo and Google Research, NeurIPS 2022) appended "Let's think step by step" and nothing else. With text-davinci-002 they report MultiArith rising from 17.7% to 78.7% and GSM8K from 10.4% to 40.7%.

The method uses two prompts. The first, "Q: [X]. A: Let's think step by step.", extracts the reasoning. The second feeds that reasoning back with a trigger such as "Therefore, the answer (arabic numerals) is" to extract the answer.

Source: Kojima et al., Table 2, accuracy in percent, as of the v4 revision of 29 January 2023. The hand-written few-shot version still beats the one-line trigger, so the phrase is a cheap baseline, not a replacement.

Why does it only work on big models, and is it really reasoning?

The authors call CoT an emergent ability of model scale: it did not help small models and only yielded gains at around 100B parameters. Smaller models produced fluent but illogical chains, which scored below standard prompting.

Their ablations rule out some cheap explanations. Writing only an equation, emitting dots to match the length of the reasoning, or giving the reasoning after the answer all performed close to baseline on the models they tested. So extra tokens alone do not explain the gain. The paper still states that CoT does not answer whether the network is actually "reasoning", which the authors leave open.

What came after: sampling, decomposition, search and training?

Four lines of work extended the idea, each with a self-reported headline number.

Lightman et al. also released PRM800K, 800,000 step-level human feedback labels. Each row measures different tasks and models, so the numbers are not comparable across rows.

The bigger shift is training. DeepSeek-R1 (arXiv 2501.12948, published in Nature, volume 645) reports that reasoning can be incentivized through pure reinforcement learning, without human-labelled reasoning trajectories. The authors describe self-reflection, verification and strategy adaptation emerging. Our DeepSeek-R1 decode covers that paper in detail. Prompting elicits step-by-step text from a model that was not built for it; reasoning models make the long chain a trained behaviour. For background on why bigger models made this possible, see GPT-3 and few-shot learners and InstructGPT and RLHF.

Does the evidence hold up?

Partly. Three problems recur.

Narrow gains. Sprague et al. (ICLR 2025) ran a meta-analysis of over 100 papers plus their own tests on 20 datasets across 14 models. CoT helped strongly on math and logic, with much smaller gains elsewhere. On MMLU, direct answering was nearly as accurate unless the question or response contained an equals sign. They also report that CoT improves symbolic execution but underperforms a dedicated symbolic solver.

Unfaithful explanations. Turpin et al. reordered multiple-choice options so the answer was always "(A)". Accuracy dropped by as much as 36% across 13 BIG-Bench Hard tasks with GPT-3.5 and Claude 1.0, and the written reasoning did not mention the cue. Lanham et al. found that as models become larger and more capable, they produce less faithful reasoning on most tasks studied, though the CoT boost did not come from extra compute alone.

Reasoning models too. Chen et al. tested reasoning models with six types of hints in the prompt. Chains of thought revealed use of a hint in at least 1% of cases where it was used, and often below 20%. Outcome-based reinforcement learning improved this at first and then plateaued.

Other limits are mechanical. Wei et al. used hand-written exemplars, so results depend on who wrote them, and the original headline numbers come from a single model family at 540B scale.

What does a working researcher take from this?

Use CoT where the task has verifiable intermediate structure, and measure it against a direct-answer baseline before paying for the extra tokens. The Sprague et al. results suggest selective use can keep accuracy while cutting inference cost.

Do not treat the written chain as evidence of how the model decided. If you need to audit behaviour, the faithfulness papers say the text can omit the real cause. Chen et al. conclude that monitoring chains of thought is promising for noticing undesired behaviour but not sufficient to rule it out.

Evaluate step-level correctness separately from final-answer accuracy. Process supervision work suggests the two can diverge, which matters if you build evals or collect reasoning data.

What's still unclear?

Whether CoT reflects reasoning or a learned imitation of reasoning format is still open; Wei et al. say so themselves. How much of the zero-shot trigger's effect is the phrase and how much is the answer-extraction step and the model's instruction tuning is not separated cleanly in the reported tables.

Faithfulness findings come from specific models and hint types. Turpin et al. tested GPT-3.5 and Claude 1.0, and newer systems may differ, though Chen et al. suggest the problem persists in reasoning models. Finally, the large gains in the self-consistency, least-to-most and Tree of Thoughts papers are lab-reported on particular benchmarks, and none of those figures should be read as general-purpose improvements.