In-context learning
Also called ICL
A model picking up a new task from examples in the prompt, without any weights changing.
Think of it like
Like showing someone three solved puzzles and watching them nail the fourth — no studying, just pattern-matching on the spot.
Example
You paste five examples of turning messy addresses into clean ones, then a sixth messy address, and the model formats it correctly.
How it actually works
The model’s weights are frozen; it’s inferring the task purely from the pattern in the context and running it during a single forward pass. This is what makes few-shot prompting work and blurs the line between “using” and “teaching” a model — though it’s temporary, lasting only as long as the examples stay in context.
For product teams
Often you can skip fine-tuning entirely — just show examples in the prompt.
For engineers
Task adaptation from prompt demonstrations at inference with no parameter updates; behavior conditioned on context.
Related
- Few-shot — Learning from a handful of examples.
- Zero-shot — Doing it with none.
- Emergent ability — An emergent capability at scale.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome