Few-shot learning is a model picking up a new task from very few examples. Classical machine learning needs thousands of examples to teach a class. Here two to five will do, because the model has already been through broad pre-training and knows the general patterns.
With language models this is a prompting matter rather than a training one. A handful of input and expected-output pairs go into the prompt, then the new input is appended. The model reads the pattern from the examples and continues in the same shape. No weights change, and the learning happens entirely inside the context window. Putting better examples in the same prompt can therefore improve results more cheaply than fine-tuning.
The choice of examples decides the outcome. If the examples cover the edge cases of the task, the model handles those cases too. If they are all easy and alike, it stumbles on hard inputs. Even ordering matters, since the last example tends to carry more weight.
A concrete case: a flow sorts incoming customer emails by urgency. One real email and its label for each level goes into the prompt. The model tags new emails against those three examples, with no separate classifier to train.
The limit is the context budget. More examples make the prompt longer, which raises cost and latency. Past a certain complexity, fine-tuning becomes the cheaper option.
