The ability of a model to perform a task it was not explicitly trained on, with no task-specific examples.