In prompting, how does zero-shot CoT differ from few-shot CoT?

## The distinction Chain-of-thought prompting means getting a model to write intermediate reasoning before committing to a final answer. There are two standard ways to elicit it, and they differ in what the prompt actually carries. **Zero-shot CoT** carries an *instruction* and nothing else — a directive such as "Let's think step by step" or "Work through the relevant factors before answering." No solved problems appear. The model decides what counts as a step, how many steps to produce, in what order, and how to present the conclusion. **Few-shot CoT** carries *demonstrations* — a small number of solved problems, each showing an input, a written-out reasoning trace, and a final answer in a consistent layout. The model infers the pattern from those examples and continues it for the new input. ## What a trigger phrase supplies A trigger supplies a mode, not knowledge. It tells the model that a longer, decomposed response is wanted rather than an immediate guess. What it cannot supply is anything specific to your problem: which considerations matter in your domain, which checks come first, what your house thresholds are, or how the answer should be laid out for downstream parsing. Two runs of the same zero-shot prompt on similar inputs can produce reasoning of very different length and structure, and the final answer may be buried in prose rather than sitting in a predictable slot. ## What worked examples supply Exemplars supply *method* and *format* simultaneously, and it is easy to forget the second half. Method: the examples encode a procedure — the order in which factors are weighed, the intermediate quantities that get named, the rule that is applied at the end. Format: they pin the output — roughly how long the reasoning runs, what sections it has, and exactly how the final answer is written, which is what makes the result reliably parseable without a separate extraction step. That second effect is the one candidates miss. Examples are not neutral illustrations; they are a strong prior over the shape of the response. ## Cost A trigger phrase is a handful of tokens. Exemplars are typically hundreds to thousands, and they are part of the input on *every* request, not a one-time setup. On a high-volume path that is a real line item, and it also consumes context that could hold retrieved material or conversation history. Caching a shared prompt prefix reduces the price of those repeated tokens but does not make them disappear, and it does not change the fact that whatever is in the exemplars is transmitted and logged on each call. ## When each is preferred Reach for **few-shot CoT** when the reasoning shape is unusual or house-specific — an internal escalation ladder, a scoring rubric, a domain notation — or when downstream code needs a strict, predictable output layout, or when zero-shot output is measurably inconsistent across similar inputs. Reach for **zero-shot CoT** when no worked solutions exist yet — a brand-new internal DSL with no solved cases is the clean example, and you simply cannot write exemplars for reasoning nobody has done before. Also prefer it when inputs vary so widely that any small set of examples would misrepresent most of them, when the token budget per call is tight relative to the task's difficulty, or when honest exemplars would have to be written from sensitive records. The two are not exclusive. A common production shape is a short instruction plus a strict output template and no full worked traces: you get format control without paying for complete exemplar reasoning. ## What has changed by mid-2026 Models trained to reason produce chains by default, so the literal phrase "let's think step by step" adds far less than it did when the technique was first reported, and on such models the effort dial has largely moved out of the prompt text and into a provider-level control. The interesting question has shifted from *how do I switch reasoning on* to *do I hand the model exemplars at all*, which is a format-and-cost decision rather than a reasoning-elicitation one. State your assumed model class when answering, because the honest answer differs between a small instruction-tuned model and a reasoning-trained frontier model. ## Common mistakes Treating few-shot as strictly better: on a task whose inputs are heterogeneous, examples can narrow the model rather than help it. Treating the trigger phrase as a universal accuracy booster: it does little on trivially easy tasks and on models that already reason. And forgetting that exemplars are re-billed and re-transmitted on every single request.