Q64

Zero-shot, one-shot, and few-shot prompting compared

PromptingZero-shotFew-shot

What's actually being asked

Define zero-shot, one-shot, and few-shot prompting, show the same task prompted all three ways, and explain which situations call for which.

Answer

The core difference

All three ask a pretrained LLM to do a task without updating its weights — the only thing that changes is how much demonstration you put inside the prompt itself (in-context learning), not any fine-tuning:

	Demonstrations in prompt	Relies on
Zero-shot	0	the model's general instruction-following ability
One-shot	1	a single example to pin down format/label vocabulary
Few-shot	2+	pattern induction from several examples

Same task, all three prompts

Task: classify a movie review's sentiment as positive or negative.

Zero-shot — only an instruction, no examples:

Classify the sentiment of this review as positive or negative.
Review: "The movie's pacing was slow but the ending was fantastic."
Sentiment:

One-shot — one worked example before the real query:

Review: "Terrible acting, I walked out halfway." → Sentiment: negative
 
Review: "The movie's pacing was slow but the ending was fantastic." → Sentiment:

Few-shot — multiple examples, ideally covering different patterns (clear positive, clear negative, and a mixed one like the target):

Review: "Terrible acting, I walked out halfway." → Sentiment: negative
Review: "Loved every second, best film this year." → Sentiment: positive
Review: "Visuals were stunning but the plot made no sense." → Sentiment: negative
 
Review: "The movie's pacing was slow but the ending was fantastic." → Sentiment:

When each is most suitable

Zero-shot: large, instruction-tuned models (RLHF-trained chat models) on common, well-understood tasks — sentiment, translation, simple summarization. Cheapest (shortest prompt, lowest latency/cost), and preferred by default when it already works, since every extra example costs context-window tokens.
One-shot: when the task itself is easy but the output format or label vocabulary is ambiguous (e.g., should the answer be "positive" or "Positive" or "4/5 stars"?) — one example disambiguates that without much extra cost.
Few-shot: niche, domain-specific, or structurally unusual tasks where the model needs to see the pattern across multiple examples to generalize correctly — e.g., extracting data into a specific JSON schema, following an idiosyncratic classification scheme, or a task the base model wasn't strongly instruction-tuned for. Few-shot also helps when examples need to show edge cases (like the mixed-sentiment review above), since one example can't cover every pattern.

Trade-off to mention: more examples = more tokens = higher cost/latency, and the model can be sensitive to the order and selection of examples (a known brittleness of in-context learning) — so few-shot isn't strictly "always better," it's a cost/robustness trade.

Exam tip

Don't confuse this axis (how many examples, in-context) with fine-tuning (which actually updates weights on many examples offline) — zero/one/few-shot all happen entirely within a single frozen-model forward pass at inference time.