Picture an assistant you show two or three worked examples to before asking your real question — and which, without ever being retrained, instantly grasps the pattern and reproduces it. That is exactly In-Context Learning (ICL): the ability of large language models to learn a task directly from the examples placed in the prompt, without any update to their weights.
Learning without retraining
Unlike fine-tuning, which actually modifies the model's parameters, ICL changes nothing. Everything happens at inference time: the provided examples (the demonstrations) condition the prediction. Three regimes are distinguished by the number of examples:
| Regime | Examples given | Typical use |
|---|---|---|
| Zero-shot | 0 | Instruction only |
| One-shot | 1 | Convey the expected format |
| Few-shot | 2 to ~50 | Stabilize a complex pattern |
How it works
The model estimates the most likely continuation conditioned on the entire supplied context:
$$ P(y \mid x, \; {(x_1, y_1), \dots, (x_k, y_k)}) $$
The example pairs act as an implicit template: the model infers the underlying rule, then applies it to the new input $x$. It is an emergent property, observed mainly at large scale.
Strengths and limits
- Advantages: no training data, instant adaptation, ideal for prototyping.
- Limits: sensitivity to the order and choice of examples, a bounded context window, less stable results than a specialized model.
In-context learning turns the prompt itself into a programming interface: you no longer reconfigure the model, you simply show it the example.