When a model performs a task correctly without being shown any examples first, relying only on its general training.
Modern LLMs are often surprisingly capable zero-shot, but performance on specialized or ambiguous tasks usually improves significantly once examples or clearer instructions are added.
