← Concept Index

Training & adaptation

In-context learning

Also called: learning from the prompt, few-shot at use time

DEFINITION

A model's ability to 'learn' a task from instructions and examples placed in the prompt, adjusting its behaviour for that conversation without any change to its weights.

WHY IT MATTERS

It's why giving examples in the prompt works so well and so immediately — and why the effect vanishes when the conversation ends. It's the cheap, reversible alternative to fine-tuning for shaping behaviour.

COMMONLY CONFUSED WITH

Fine-tuning or permanent learning. In-context learning lasts only for that context window; the model hasn't actually been changed.