Definition
Zero-shot learning has two related uses that should be named explicitly.
In classical machine learning, it means predicting classes that had no labeled training examples by using shared semantic information such as attributes or class descriptions. Mark Palatucci and coauthors' 2009 paper, Zero-shot Learning with Semantic Output Codes, is one of the earliest documented uses of the modern term and helped establish this unseen-class meaning.
In current language-model work, zero-shot prompting usually means asking a model to perform a task without demonstrations in the prompt. The model still brings knowledge from pretraining, instruction tuning, and any tools or retrieved context. "Zero-shot" therefore does not mean the system has never encountered related concepts, examples, or instructions.
GPT-3's 2020 evaluation and FLAN's 2021 instruction-tuning work helped popularize the current LLM sense. They did not coin the broader learning concept.
Distinguish it from nearby terms
- Few-shot prompting supplies several demonstrations in context. One-shot prompting supplies one.
- Zero-shot prompting describes the prompt configuration. Zero-shot learning in the classical sense describes generalization to unseen output classes.
- A held-out task can still be evaluated with demonstrations. "Held out" describes separation from development, not the number of examples in a prompt.
Check your understanding
A model receives no examples in its prompt but was fine-tuned on the benchmark's training split. Is the run zero-shot prompting, zero-shot learning, both, or neither?