Zero-shot chain-of-thought is a prompting technique in which a generic reasoning trigger, most famously the phrase “Let’s think step by step”, is appended to a question so that a language model writes out intermediate reasoning before its answer, without any worked examples in the prompt. In prompting, it is the simplest way to elicit chain-of-thought reasoning, and it is often abbreviated as zero-shot CoT.
The technique was introduced in 2022 in research showing that large language models are zero-shot reasoners: a single instruction, rather than carefully written demonstrations, was enough to raise accuracy sharply on arithmetic, symbolic and commonsense reasoning benchmarks.
How the reasoning trigger works
Zero-shot chain-of-thought works by changing what the model generates first. A model asked a question directly tends to emit an answer immediately, while a model prompted with a reasoning trigger generates a sequence of intermediate steps, and each step becomes context that conditions the next token.
Because a transformer performs a fixed amount of computation per token, writing out steps gives the model more serial computation and a visible scratchpad. Problems that require several dependent operations, such as multi-step arithmetic, benefit most, since each partial result is stored in the text rather than computed implicitly.
The original method uses two prompting stages, summarized below.
| Stage | Prompt addition | Purpose |
|---|---|---|
| Reasoning extraction | Question followed by the reasoning trigger | Generate a free-form chain of reasoning |
| Answer extraction | Question, reasoning and an answer cue | Convert the reasoning into a clean final answer |
The answer cue, for example “Therefore, the answer (arabic numerals) is”, forces a parseable final answer. Modern applications often merge both stages into one call by asking for the reasoning followed by a clearly labeled answer line.
Trigger phrases and their effect
The wording of the trigger matters, because models respond to instructive phrases far more than to neutral ones. Early experiments compared many templates and found that phrases encouraging explicit reasoning outperformed misleading or irrelevant text by a wide margin.
Common triggers and their typical role are compared in the following table.
| Trigger phrase | Origin | Typical effect |
|---|---|---|
| “Let’s think step by step” | Original zero-shot CoT study | Strong baseline for arithmetic and logic |
| Work it out step by step to be sure the answer is right | Automatic prompt search | Slightly higher accuracy in some tests |
| First understand the problem and devise a plan | Plan-and-solve prompting | Fewer skipped steps in long problems |
| “Take a deep breath and work on this problem step by step” | Optimizer-generated prompts | Gains on specific models only |
Gains from a particular phrasing are model-specific and rarely transfer exactly. Teams should treat any trigger as a hypothesis and confirm it with prompt evaluation on their own data.
How it compares with related prompting methods
Zero-shot chain-of-thought sits at the low-effort end of reasoning techniques. It needs no examples and no extra infrastructure, while related methods trade more prompt work or more model calls for higher accuracy.
| Technique | Examples needed | Model calls | Main advantage |
|---|---|---|---|
| Zero-shot chain-of-thought | None | One or two | Minimal setup |
| Few-shot chain-of-thought | Several worked examples | One | Controls reasoning format and style |
| Auto-CoT | Generated automatically | Many during setup | Few-shot quality without manual writing |
| Self-consistency | Optional | Five to forty samples | Majority vote reduces random errors |
| Plan-and-solve prompting | None | One | Explicit planning before calculation |
In practice these methods combine well. Self-consistency, for example, is often applied on top of zero-shot CoT: the same triggered prompt is sampled several times and the most frequent final answer is selected.
Few-shot chain-of-thought generally remains more accurate on domain-specific tasks, because demonstrations show the exact reasoning style and answer format expected. Zero-shot CoT is therefore best seen as a strong default baseline: if it already meets the accuracy target, writing and maintaining examples adds cost without clear benefit.
When to use zero-shot chain-of-thought
Zero-shot chain-of-thought is most useful for multi-step problems with a verifiable answer, such as word problems, logic puzzles, date calculations and structured decisions. It adds little to simple factual lookups or short classification, where extra text mainly increases latency and cost.
| Situation | Recommendation | Reason |
|---|---|---|
| Multi-step arithmetic or logic | Use the trigger | Intermediate steps reduce skipped operations |
| Simple fact retrieval | Skip the trigger | No reasoning chain to improve |
| Strict output format needed | Use trigger plus answer cue | Separates reasoning from the parsed result |
| Dedicated reasoning model | Usually skip the trigger | Reasoning is already built in |
| High-stakes answers | Combine with self-consistency or verification | Single chains can be confidently wrong |
Limitations and common misconceptions
Zero-shot chain-of-thought improves average accuracy but does not guarantee correct reasoning. Several limitations are well documented.
- Scale dependence: the original gains appeared mainly in large models, and small models often produced fluent but wrong chains.
- Unfaithful explanations: the written reasoning is not necessarily the process that produced the answer, so it should not be read as a reliable audit trail.
- Error propagation: one early mistake carries through every later step, often with confident wording.
- Diminishing returns: instruction-tuned and reasoning models already reason by default, so the explicit trigger adds less than it did in 2022.
- Higher token usage: reasoning text can multiply output length several times, which matters at high request volumes.
A common misconception is that the phrase itself carries special power. The effect comes from shifting the model toward explicit intermediate computation, and any clear instruction with the same intent tends to work similarly.
The idea in brief
Zero-shot chain-of-thought means asking an AI model to show its work before answering, without giving it any solved examples. Adding a short phrase such as “think step by step” makes the model write down each stage of its reasoning.
It resembles a teacher asking a student to solve a math problem on paper rather than in their head. Writing each step makes it harder to skip a calculation, although a careful student can still make a mistake on paper.
A practical scenario
A logistics company uses a language model to answer staff questions about delivery schedules, such as how many working days remain before a shipment arrives when a public holiday falls in between. Direct prompts frequently miscount weekends and holidays.
The team appends a reasoning instruction to every question and asks the model to finish with a single line labeled with the final number. On a test set of 200 historical questions with known answers, accuracy rises from roughly two thirds to above 90 percent, while average response length grows about fourfold.
Because the longer outputs raise costs, the team applies the reasoning trigger only to date and quantity questions detected by a simple classifier, and it keeps short direct prompts for routine lookups such as office addresses.