What is few-shot prompting? Few-shot vs zero-shot

Also called few-shot learning, in-context learning, k-shot prompting, multishot prompting

Definition

Few-shot prompting is a technique that puts a few example inputs paired with desired outputs in a prompt, so a language model infers the task, format, and labels without changing its weights.

Few-shot prompting is the practice of placing a few worked examples, each an input paired with the desired output, in a prompt before the real input. The model infers the task, the output format, and the label set from the examples and continues the pattern, with no training or weight updates. It is one of the cheapest ways to make a large language model produce consistent output, and the examples often live in the system prompt.

How few-shot prompting works#

Few-shot prompting works through in-context learning: the model conditions on the demonstrations in its input and applies the same pattern to the new case. The GPT-3 paper (Brown et al., 2020) named the settings by example count: zero-shot gives only an instruction, one-shot adds one demonstration, and few-shot adds several.

For GPT-3, few-shot meant roughly 10 to 100 examples, as many as fit in its 2,048-token context window. The paper found that scale improves this ability. Its 175-billion-parameter model performed many tasks from demonstrations alone, sometimes matching earlier fine-tuned systems.

On classification and multiple-choice tasks, the examples teach format more than correctness. Min et al. (2022) replaced the labels in demonstrations with random ones and found that performance barely dropped, across 12 models including GPT-3. What mattered was that the examples showed the label space, the distribution of input text, and the overall format. Choose examples that resemble real inputs and cover the full range of expected outputs.

Few-shot vs zero-shot prompting#

Zero-shot prompting gives the model only an instruction; few-shot prompting adds worked examples. Examples are the more reliable way to fix an output format, a tone, or a custom label set. They also cost input tokens on every request and can over-steer, since a model may copy surface features of the examples, such as their length or phrasing.

Zero-shot Few-shot Fine-tuning
What you provide An instruction An instruction plus examples A training dataset
Weights change No No Yes
Prompt length Shortest Grows with each example Can shrink
Good fit Clear, common tasks Strict formats, custom labels Behavior too complex to show in a prompt

Anthropic's prompting best practices recommend 3 to 5 examples that are relevant and diverse, wrapped in <example> tags so the model can tell them apart from instructions. A fixed block of examples sits at the same position in every request, which makes it a good candidate for prompt caching.

Reasoning models change the default. OpenAI's reasoning best practices advise trying zero-shot first and adding examples only if needed, matched closely to the instructions. Anthropic notes that worked examples still shape how its models approach similar problems in their thinking.

The LLM benchmarks guide lists prompt format, including few-shot examples, among the main reasons published scores disagree.

Cite this definition

The AI Engineering Editors. "What is few-shot prompting? Few-shot vs zero-shot." The AI Engineering, updated Nov 06, 2026. https://theaiengineering.com/glossary/few-shot-prompting

Title
What is few-shot prompting? Few-shot vs zero-shot
Author
The AI Engineering Editors
Publisher
The AI Engineering
Updated
View as Markdown

Part of the Prompt and context engineering hub · Updated