ModelRefs / Few-Shot Prompting — AI Glossary

Few-Shot Prompting — AI Glossary

Providing a small number of example input-output pairs in the prompt to demonstrate the desired behavior to the model.

Overview

Few-shot prompting works because LLMs perform in-context learning. 3–8 examples is typical; more rarely helps and consumes context. Example selection matters — diverse, representative examples outperform random sampling.

Reference details

Topic	prompting
Last reviewed	2026-06-24

How it works

The examples are concatenated into the prompt as ordinary text in a repeated format, so from the model's side they are context rather than a distinct mechanism.
During the forward pass attention over those examples induces the input-to-output mapping they demonstrate, which is why the behaviour appears without any weight change.
The format carries a large share of the signal: consistent delimiters and a consistent field order tell the model where its own output begins, independently of what the examples contain.
The distribution of labels across the examples biases the output distribution, so an unbalanced set shifts predictions toward whichever label appears most.
Position matters as well, with examples nearest the query typically weighted most heavily, which is why ordering is a parameter rather than an arbitrary choice.
Selection can be fixed or retrieved per query from a larger pool by similarity to the input, and the retrieved form generally outperforms a static set because the demonstrated edge cases are closer to the case at hand.

Example: Examples teach the edge cases, not the task

For sentiment labelling, the useful examples are not three clear positives. They are the ones that pin down your conventions: sarcasm, mixed sentiment, and a review that is about delivery rather than the product. Examples are where you encode the decisions a definition cannot express.

Commonly confused with

In-context learning does not update the model. The examples influence one response and are gone; nothing is retained between calls. That is the difference from fine-tuning, and the reason few-shot costs tokens on every single request.

When to use it

Reach for it when:

The desired output form is easier to show than to describe
There are conventions and edge cases a definition would not capture
You need behaviour changeable without retraining

Reach for something else when:

Examples would consume context you need for actual content
The pattern is stable and high-volume — fine-tuning amortises better
Your examples are unrepresentative; the model will copy the skew faithfully

Referenced by

This term is used by the following ModelRefs references:

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Few-Shot Prompting — AI Glossary.

Frequently asked questions

What is Few-Shot Prompting?