Zero-shot describes the ability of a language model to tackle a task without having received specific examples of it in the prompt or dedicated training. The model relies solely on the general knowledge acquired during its pre-training and on the instructions it receives in natural language. For example, if you ask to "classify this email as spam or legitimate" without showing any previously solved cases, you are using a zero-shot approach.
Its importance lies in its flexibility: it allows for solving new problems without the need to collect labeled data or retrain, which saves time and resources. It is the most direct way to interact with a model and the foundation upon which more elaborate techniques are built.
It is useful to distinguish it from other related approaches:
- Zero-shot: without examples in the prompt.
- Few-shot: a few examples are included to guide the model.
In practice, zero-shot works well for common tasks, but its accuracy may drop in very specific problems or those requiring a particular output format, where adding examples usually improves results.