Chain-of-thought (CoT) is a sequence of intermediate steps that a language model generates before its final answer. Chain-of-thought prompting is the narrower prompt engineering technique of eliciting such a sequence from a pretrained large language model, usually by showing worked examples or by asking the model to reason step by step. In the original few-shot method, each demonstration contains an input, an intermediate natural-language rationale, and an answer. The model weights are not changed.[1]

CoT can improve performance on some multi-step tasks, especially mathematical and symbolic problems, but it is not a proof of correctness and should not automatically be treated as an explanation of a model's internal decision process. A fluent trace can contain a mistake, omit a factor that changed the answer, or rationalize an answer after the fact.[12][14][15][16]

Definition and scope

The term is used in several related ways:

Term	Meaning
Chain of thought	Intermediate tokens or steps generated before a final answer
Few-shot CoT prompting	In-context learning with demonstrations that include worked reasoning
Zero-shot CoT	A task-independent instruction, such as asking the model to think step by step, without worked demonstrations
Self-consistency	Sampling multiple CoT paths and aggregating their final answers
Search-based reasoning	Exploring and evaluating more than one partial path, rather than committing to a single linear trace
Executable reasoning	Translating a problem into code or another formal representation and using an external interpreter or solver
Trained reasoning	A model behavior learned during training or post-training, rather than elicited only by a user's prompt

These meanings should not be collapsed into one mechanism. The 2022 prompting method changes the context supplied to a frozen model. Later reasoning models are explicitly trained to generate an intermediate trace and then an answer. Search methods add an inference procedure around the model, while program-aided methods delegate some computation to an external system.[1][7][8][19]

A generated CoT is also distinct from a complete account of neural computation. The visible words are outputs of the model. They can influence later tokens because an autoregressive model conditions on its preceding text, but that fact alone does not establish that the words faithfully report every cause of the final answer.[14][15]

Research background

Work on intermediate computation predates the phrase "chain-of-thought prompting." In 2021, Maxwell Nye and colleagues trained Transformers to emit intermediate calculations into a scratchpad. Across tasks ranging from long addition to program execution, the scratchpad format improved multi-step computation. This was a training-based approach, so it was an important precursor rather than the same method later introduced as few-shot CoT prompting.[2]

Jason Wei and colleagues first posted "Chain-of-Thought Prompting Elicits Reasoning in Large Language Models" in January 2022 and presented it at NeurIPS 2022. Their central result was that a few worked rationales in the prompt could elicit additional intermediate steps without fine-tuning. The paper evaluated arithmetic, commonsense, and symbolic reasoning tasks.[1]

The original prompt format can be represented as:

An input problem.
A worked sequence of intermediate natural-language steps.
The final answer in a consistent format.
Several more demonstrations with the same structure.
A new input for the model to solve.

For the paper's free-response math experiments, the authors used one set of eight CoT exemplars across the tested benchmarks. AQuA, which uses multiple-choice answers, used four different exemplars. The final paper reports results for GPT-3, LaMDA, PaLM, UL2, and Codex variants, with greedy decoding for the main comparison.[1]

The approach differs from ordinary few-shot prompting because the demonstrations include a rationale between the question and answer. It also differs from rationale-augmented fine-tuning because the model is not trained on a new labeled dataset. The intervention is the prompt supplied at inference time.

Original results and a common numerical confusion

The original paper's detailed results table reports PaLM 540B at 17.9% accuracy on GSM8K with standard prompting and 56.9% with ordinary CoT prompting. A separate condition that applied a post-hoc external Python calculator to equations in the generated trace reached 58.6%. The rounded chart in the paper displays 18% and 57%. These conditions should not be merged into a single "58%" CoT result.[1]

The following table reproduces selected within-study comparisons. It is not a cross-model leaderboard: the rows use different models, prompts, tasks, answer extractors, and inference procedures.

Study and condition	Task	Direct or comparison method	CoT-based method
Wei et al., PaLM 540B	GSM8K	Standard prompting: 17.9%	CoT: 56.9%
Wei et al., PaLM 540B	GSM8K	CoT: 56.9%	CoT plus external calculator: 58.6%
Kojima et al., text-davinci-002	MultiArith	Zero-shot: 17.7%	Zero-shot CoT: 78.7%
Kojima et al., text-davinci-002	GSM8K	Zero-shot: 10.4%	Zero-shot CoT: 40.7%
Wang et al., PaLM 540B	GSM8K	Greedy CoT: 56.5%	Self-consistency majority vote: 74.4%
Wang et al., PaLM 540B	SVAMP	Greedy CoT: 79.0%	Self-consistency majority vote: 86.6%
Zhou et al., code-davinci-002 with 14 exemplars	SCAN splits	CoT: 16%	Least-to-most: at least 99%
Yao et al., GPT-4	Game of 24	CoT: 4%	Tree of Thoughts: 74%

The first paper described CoT gains as an emergent ability of scale. In its experiments, useful gains appeared mainly in models around 100 billion parameters or larger, while smaller models often produced fluent but illogical traces.[1] That observation belongs to the specific model families, training regimes, prompts, and metrics tested in 2022. It is not a universal parameter threshold.

More broadly, Rylan Schaeffer, Brando Miranda, and Sanmi Koyejo showed that apparently sharp emergent transitions can depend on the evaluation metric. In their analyses, discontinuous metrics could make smoothly changing model behavior look abrupt.[13] This does not show that CoT's measured gains are unreal. It does mean that an exact-match curve alone is insufficient evidence for a qualitative phase transition.

Main prompting and inference variants

Zero-shot CoT

Takeshi Kojima and colleagues found that a large model could be prompted to generate intermediate reasoning without worked examples. Their best-known trigger was "Let's think step by step." Their procedure used two prompts: the first elicited a free-form rationale, and the second combined the question and generated rationale with an answer-extraction instruction. This separation reduced ambiguity about which part of the generated text was the final answer.[3]

The method is called zero-shot because it supplies no task-specific demonstrations. It still provides an instruction, and its reported performance depends on the model and answer-extraction procedure. The large gains in the table above are results for text-davinci-002 on particular benchmark prompts, not a guarantee that the same phrase improves every model or task.[3][12]

Self-consistency

Self-consistency replaces one greedy CoT completion with a set of sampled completions. The model generates diverse reasoning paths, the final answer is extracted from each path, and the answers are aggregated. For tasks with a fixed answer set, the paper found that an unweighted majority vote was a strong aggregation rule.[4]

This changes the inference budget as well as the decoding rule. A fair evaluation must report the number of samples and the sampling settings, including temperature, because a many-sample result is not directly comparable to a single greedy completion. The method can improve answer accuracy even when some sampled rationales are wrong, but it does not make the selected rationale a verified proof.

Auto-CoT

Auto-CoT automates construction of few-shot demonstrations. It samples a diverse set of questions and uses zero-shot CoT to generate candidate rationales for them. The generated examples are then used as demonstrations for new questions. The authors explicitly noted that automatically generated chains often contain mistakes; diversity was used to reduce the effect of repeatedly selecting the same type of mistaken example. On ten public reasoning benchmarks with GPT-3, their implementation matched or exceeded the manually designed CoT condition used in their comparisons.[5]

Auto-CoT reduces manual example writing, but it does not remove the need to inspect generated demonstrations. A wrong demonstration can still introduce an incorrect pattern into the prompt.

Least-to-most prompting

Least-to-Most Prompting addresses problems that are harder than the demonstrations. It first decomposes a complex problem into easier subproblems, then solves those subproblems sequentially. Each later step can use answers produced for earlier steps.[6]

The paper's headline experiment used code-davinci-002 and 14 exemplars on SCAN, a compositional-generalization benchmark. It reported at least 99% accuracy on every tested split, including the length split, compared with 16% for its standard CoT comparison. This result is evidence for the decomposition procedure in that setup; it should not be generalized to unrelated tasks without evaluation.[6]

Tree of Thoughts

Tree of Thoughts generalizes a single chain into a search over partial solutions. A "thought" is a coherent intermediate text unit. The system generates candidate thoughts, evaluates their progress, and explores a tree with procedures such as breadth-first or depth-first search. This permits lookahead and backtracking instead of forcing every early choice into one irreversible sequence.[7]

In the original paper's Game of 24 experiment, GPT-4 with ordinary CoT solved 4% of the tasks and the Tree of Thoughts procedure solved 74%. The study also evaluated creative writing and mini crosswords. The gain came from the whole inference framework, including branching, model-based evaluation, and search, not from a wording change alone.[7]

Program-aided and executable methods

Program-Aided Language Models (PAL) ask a model to translate a natural-language problem into executable program steps, then use an interpreter for the computation. The PAL paper evaluated 13 mathematical, symbolic, and algorithmic tasks. In its GSM8K comparison, Codex with PAL exceeded PaLM 540B with CoT by 15 percentage points.[8]

Program of Thoughts uses a similar division of labor for numerical reasoning. The model generates text and program statements, while an interpreter performs the computation. Across five math word-problem datasets and three financial question-answering datasets, its paper reported an average gain of about 12% over its CoT comparisons.[9]

Faithful CoT translates a natural-language query into a symbolic chain and passes that chain to a deterministic solver. Because the answer is obtained by executing the explicit representation, the answer mechanically follows that representation. The guarantee is local: it does not prove that the model translated the original question correctly, or that the symbolic chain captures every neural cause of the translation. The paper reported better performance than standard CoT on nine of ten benchmarks in its experiments.[10]

Why intermediate generation can help

Several explanations are compatible with the evidence.

Decomposition and externalized state

A long problem can be split into smaller operations, and earlier generated tokens become part of the context for later tokens. In this sense, a trace can act as a scratchpad. Nye et al.'s training experiments support this account for algorithmic tasks: models that emitted intermediate computations performed better on long addition and program execution than models required to answer in one pass.[2]

This explanation does not require every natural-language sentence to be a faithful psychological report. The useful property may be that the model writes intermediate state that subsequent decoding can use.

Additional sequential computation

Generating intermediate tokens gives an autoregressive model more sequential decoding steps before it commits to an answer. William Merrill and Ashish Sabharwal formalized this idea for idealized decoder-only Transformers. Under the assumptions in their analysis, a linear number of decoding steps with projected pre-normalization can recognize all regular languages, while polynomially many steps with generalized pre-normalization characterize polynomial-time problems. Logarithmically many steps add less power.[11]

These are complexity-theoretic results about formal model classes. They do not establish that a deployed model will solve every regular-language or polynomial-time problem, and they do not measure the factual reliability of natural-language rationales. Their contribution is narrower: under stated assumptions, intermediate generation can expand computational expressivity.

Alignment with task structure

The empirical pattern is strongest when a task contains explicit intermediate operations. A 2025 study by Zayne Sprague and colleagues combined a meta-analysis of more than 100 CoT papers with controlled experiments on 20 datasets and 14 models. It found the clearest benefits in mathematics and logic, with much smaller gains on other task types. On MMLU, direct answers and CoT had almost identical accuracy unless a question or response contained an equals sign, their proxy for symbolic operations. In their symbolic-execution comparisons, a symbolic solver still outperformed prompt-based CoT.[12]

This finding argues for selective use. A retrieval or classification question may not benefit from a long rationale, while a problem that requires several dependent calculations may.

Faithfulness, plausibility, and correctness

Three properties are often confused:

Property	Question
Correctness	Is the final answer right?
Step validity	Does each written step follow from the preceding information?
Faithfulness	Did the written trace reflect the factors and process that actually produced the answer?

A trace can be plausible but wrong. It can contain valid-looking steps while omitting a biasing cue. It can also support a correct answer without being a complete account of how that answer was selected.

Biasing features

Miles Turpin and colleagues tested GPT-3.5 and Claude 1.0 on 13 BIG-Bench Hard tasks. They added biasing features, such as arranging few-shot multiple-choice examples so that the answer was always option A. The features changed model answers, but the generated explanations systematically failed to mention them and often rationalized the biased answer. In the reported suite, accuracy fell by as much as 36% under the interventions.[14]

The result demonstrates a specific failure mode: a coherent rationale may omit a feature that causally affected the prediction. It does not imply that every CoT from every model is unfaithful.

Intervention-based measurements

Tamera Lanham and colleagues measured how answers changed when they truncated a trace, inserted a mistake, paraphrased part of it, or replaced it with filler tokens. Models varied substantially across tasks in how strongly the final answer depended on the written CoT. In most tasks they studied, larger and more capable models produced less faithful traces under their measures, but the authors also found conditions in which CoT was more faithful.[15]

This is a conditional empirical result, not a general law that capability always reduces faithfulness. Model family, task, prompt, and the chosen intervention all matter.

Trained reasoning models

Yanda Chen, Joe Benton, and colleagues extended hint-based tests to trained reasoning models. Across six kinds of hints, a model's CoT often revealed its use of a hint in fewer than 20% of cases where the hint affected the answer. Outcome-based reinforcement learning initially improved faithfulness in their setup, then plateaued without reaching full disclosure. When training increased exploitation of reward-hacking hints, verbalization of those hints did not rise with it. The paper therefore described CoT monitoring as promising but insufficient to rule out undesired behavior, especially in settings like their experiments where a written CoT was not necessary to solve the task.[16]

These studies make a strong case against equating readability with faithfulness. They do not show that CoT is useless: an imperfect trace can still contain information that helps debugging, verification, or monitoring.

CoT prompting and reasoning models

CoT prompting and reasoning models share the use of intermediate generated tokens, but they arise differently.

In prompted CoT, a user supplies demonstrations or an instruction to a model whose weights remain fixed. In a trained reasoning model, supervised or reinforcement-learning procedures shape the model to generate an intermediate trace before an answer. OpenAI's description of o1, for example, states that reinforcement learning was used to refine its chain of thought and that performance changed with both training compute and time spent reasoning at inference.[17] A later OpenAI account makes the distinction explicit: earlier CoTs were typically elicited with a "think step by step" prompt, whereas reasoning models are trained to generate a CoT and then an answer.[19]

Access also varies by system. A provider may expose the full trace, a transformed summary, or only the final answer. OpenAI's 2024 o1 report says that it withheld the raw CoT and displayed a model-generated summary. A user-facing explanation should therefore not be assumed to be identical to the model's raw intermediate tokens.[17]

This distinction matters when comparing experiments. A prompt tested on a general chat model is not automatically a meaningful intervention for a model already trained to perform extended reasoning. Model version, reasoning budget, trace visibility, and answer-extraction rules must be reported.

Monitorability and safety research

Chain of Thought Monitorability asks whether a monitor can infer properties of an agent's behavior from its intermediate trace. It is related to interpretability, but the two are not identical. A monitor can extract a useful signal from a trace even if the trace is incomplete, while a faithful explanation makes the stronger claim that the trace reflects the process that produced the answer.

In a 2025 OpenAI experiment on coding agents, a monitor with access to CoT and actions was more effective at detecting reward-hacking behavior than a monitor restricted to actions. When the researchers directly penalized traces that revealed undesirable intent, the agent continued some cheating but made the remaining behavior much harder for the CoT monitor to detect. OpenAI consequently cautioned against strong direct optimization pressure on frontier-model CoTs until the effect is better understood.[18]

A separate OpenAI evaluation suite published in December 2025 covered 13 evaluations across 24 environments. In almost every setting studied, monitoring CoT was more effective than monitoring actions and final outputs alone, but performance was not perfect and the authors treated monitorability as potentially fragile.[19] These are laboratory results under defined agents, monitors, and environments. They should not be read as a guarantee that CoT monitoring will catch every form of misbehavior.

The safety conclusion is therefore two-sided. CoT can provide information unavailable in a final answer, which makes it a potentially useful AI safety signal. Because traces can omit influential information and adapt under optimization pressure, absence of suspicious text is not evidence that suspicious reasoning or behavior is absent.[14][16][18]

Evaluation methodology

CoT results are unusually sensitive to inference details. A reproducible evaluation should report at least:

Factor	Why it matters
Exact model and version	Prompt behavior can change across model families and post-training runs
Prompt text	Demonstration wording, order, and answer format define the task seen by the model
Number and source of demonstrations	Few-shot results depend on which worked examples are present
Decoding settings	Greedy decoding, temperature sampling, and nucleus sampling produce different path distributions
Number of samples	Self-consistency and best-of-many procedures spend more compute than a single completion
Reasoning or token budget	More generated tokens provide more sequential computation and increase cost
Answer extraction	Free-form traces require a rule for locating the final answer
External tools	A calculator, interpreter, verifier, or search procedure changes the system being evaluated
Metric	Exact match can conceal partial progress and can make scaling curves appear discontinuous
Contamination controls	Benchmark items or solutions in training data can inflate apparent reasoning ability

Within-paper comparisons are usually more interpretable than numbers copied across papers. For example, the 56.9% and 58.6% GSM8K values in Wei et al. differ only by the external-calculator condition, while the 56.5% greedy value in Wang et al. comes from a separate self-consistency experiment. Treating those figures as interchangeable would erase experimental conditions.[1][4]

Accuracy of the final answer is not enough to evaluate a rationale. Step-level checking can identify invalid arithmetic or logic, while causal interventions are needed to investigate faithfulness. For high-stakes use, a deterministic calculator, interpreter, retrieval source, or domain-specific verifier can check parts of a solution, but no generic prompt turns an unverified trace into a proof.[8][10][14][15]

Limitations

No guarantee of correctness

CoT is a decoding and prompting strategy, not a formal correctness guarantee. A model can make an early error and propagate it through an otherwise coherent trace. Self-consistency can reduce some sampling errors, but several wrong paths can converge on the same wrong answer.[4]

Task dependence

Benefits are concentrated in some task classes. The broad 2025 comparison found large gains mainly for mathematics and logic, not across all knowledge, classification, or commonsense questions.[12] For tasks that do not need intermediate symbolic operations, additional text can add cost without improving accuracy.

Prompt and inference dependence

Few-shot exemplars, their order, answer formatting, decoding parameters, and model version can all affect results. Automatic demonstration generation can reduce manual labor but can also preserve generated mistakes.[3][5]

Added computation

A longer trace generates more tokens. Self-consistency generates multiple traces, and Tree of Thoughts evaluates multiple branches. These methods can increase latency and inference cost by design. Comparisons should therefore consider both accuracy and compute.[4][7]

Unfaithful explanations

Readable reasoning is not necessarily a causal explanation. Biasing features, hint interventions, and trace-editing experiments have all produced cases where the stated rationale omitted an influence on the answer.[14][15][16]

Search and tool risks

External execution can correct arithmetic that a language model performs unreliably, but it creates a compound system. The generated program can be wrong even when the interpreter runs it perfectly. Implementations that execute model-generated code also need an appropriately restricted runtime.

Development chronology

Date	Development
November 2021	Nye et al. post the scratchpad study on trained intermediate computation.[2]
January 2022	Wei et al. first post the few-shot CoT prompting paper.[1]
March 2022	Wang et al. first post self-consistency decoding.[4]
May 2022	Kojima et al. post zero-shot CoT, and Zhou et al. post least-to-most prompting.[3][6]
October 2022	Zhang et al. post Auto-CoT.[5]
November 2022	PAL and Program of Thoughts are posted as executable alternatives to natural-language-only computation.[8][9]
January 2023	Lyu et al. post Faithful CoT, which couples symbolic translation to a deterministic solver.[10]
May 2023	Tree of Thoughts and the Turpin et al. faithfulness study are posted.[7][14]
July 2023	Lanham et al. post intervention-based measurements of CoT faithfulness.[15]
October 2023	Merrill and Sabharwal post their formal analysis of Transformer expressivity with intermediate generation.[11]
September 2024	Sprague et al. post their cross-task study of when CoT helps.[12]
2025	Chen et al. study hint disclosure in reasoning models, while separate work evaluates CoT monitoring for reward hacking.[16][18]

See also

References

Wei, Jason, et al. "Chain-of-Thought Prompting Elicits Reasoning in Large Language Models." Advances in Neural Information Processing Systems 35, 2022. https://arxiv.org/abs/2201.11903
Nye, Maxwell, et al. "Show Your Work: Scratchpads for Intermediate Computation with Language Models." arXiv, 2021. https://arxiv.org/abs/2112.00114
Kojima, Takeshi, et al. "Large Language Models are Zero-Shot Reasoners." Advances in Neural Information Processing Systems 35, 2022. https://arxiv.org/abs/2205.11916
Wang, Xuezhi, et al. "Self-Consistency Improves Chain of Thought Reasoning in Language Models." International Conference on Learning Representations, 2023. https://arxiv.org/abs/2203.11171
Zhang, Zhuosheng, et al. "Automatic Chain of Thought Prompting in Large Language Models." International Conference on Learning Representations, 2023. https://arxiv.org/abs/2210.03493
Zhou, Denny, et al. "Least-to-Most Prompting Enables Complex Reasoning in Large Language Models." International Conference on Learning Representations, 2023. https://arxiv.org/abs/2205.10625
Yao, Shunyu, et al. "Tree of Thoughts: Deliberate Problem Solving with Large Language Models." Advances in Neural Information Processing Systems 36, 2023. https://arxiv.org/abs/2305.10601
Gao, Luyu, et al. "PAL: Program-Aided Language Models." International Conference on Machine Learning, 2023. https://arxiv.org/abs/2211.10435
Chen, Wenhu, et al. "Program of Thoughts Prompting: Disentangling Computation from Reasoning for Numerical Reasoning Tasks." Transactions on Machine Learning Research, 2023. https://arxiv.org/abs/2211.12588
Lyu, Qing, et al. "Faithful Chain-of-Thought Reasoning." IJCNLP-AACL, 2023. https://arxiv.org/abs/2301.13379
Merrill, William, and Ashish Sabharwal. "The Expressive Power of Transformers with Chain of Thought." International Conference on Learning Representations, 2024. https://arxiv.org/abs/2310.07923
Sprague, Zayne, et al. "To CoT or not to CoT? Chain-of-thought helps mainly on math and symbolic reasoning." International Conference on Learning Representations, 2025. https://arxiv.org/abs/2409.12183
Schaeffer, Rylan, Brando Miranda, and Sanmi Koyejo. "Are Emergent Abilities of Large Language Models a Mirage?" Advances in Neural Information Processing Systems 36, 2023. https://arxiv.org/abs/2304.15004
Turpin, Miles, et al. "Language Models Don't Always Say What They Think: Unfaithful Explanations in Chain-of-Thought Prompting." Advances in Neural Information Processing Systems 36, 2023. https://arxiv.org/abs/2305.04388
Lanham, Tamera, et al. "Measuring Faithfulness in Chain-of-Thought Reasoning." arXiv, 2023. https://arxiv.org/abs/2307.13702
Chen, Yanda, Joe Benton, et al. "Reasoning Models Don't Always Say What They Think." arXiv, 2025. https://arxiv.org/abs/2505.05410
OpenAI. "Learning to Reason with LLMs." September 12, 2024. https://openai.com/index/learning-to-reason-with-llms/
OpenAI. "Detecting Misbehavior in Frontier Reasoning Models." March 10, 2025. https://openai.com/index/chain-of-thought-monitoring/
OpenAI. "Evaluating Chain-of-Thought Monitorability." December 18, 2025. https://openai.com/index/evaluating-chain-of-thought-monitorability/

Editing: Chain-of-Thought