Zero-shot refers to performing a task without any task-specific examples in the prompt, relying entirely on knowledge from pre-training and instruction tuning. A model answering a question about a topic it was never explicitly trained on is zero-shot. Zero-shot performance has improved dramatically with scale and instruction tuning, and it's a key measure of a model's generalization ability. The term originates from zero-shot learning in computer vision, where classifiers recognize categories they've never seen labeled examples of.