What is Zero-Shot Learning?

In classical machine learning, a model needed labeled training examples for every category it was expected to recognize, so classifying a brand-new category meant collecting fresh data and retraining. Large language models broke this constraint: because they're trained on such broad, diverse text, they can often perform a novel task described only in a prompt, such as classifying customer messages into categories the model has never explicitly been trained on, without any task-specific fine-tuning or example data at all. Zero-shot performance sits on a spectrum with few-shot prompting, where a handful of labeled examples are included directly in the prompt to steer the model's behavior toward exactly the desired output format or edge-case handling. Zero-shot is fastest to deploy since it needs no example curation, but few-shot generally produces more consistent, precisely-formatted results for tasks with subtle distinctions the model might otherwise misjudge on its own.