Zero-shot learning is a model's ability to perform a task without task-specific training examples for that exact task.
In modern language models, zero-shot behavior often depends on broad pretraining and clear instructions in the prompt.