Zero-Shot
Getting a model to do a task from instructions alone, with no examples in the prompt.
Zero-shot means asking a model to perform a task purely from a description, with no worked examples provided — relying on what it learned in pretraining. Adding a few examples in the prompt makes it few-shot instead.
Worked example: “Classify the sentiment of this review: …” with no example labels is zero-shot; prepend two labeled reviews and it becomes two-shot, which often lifts accuracy on harder or unusual formats. Gotcha: zero-shot leans entirely on the instruction being clear and the task being well-represented in pretraining; for niche formats or strict output shapes, a few examples usually beats a longer instruction.