What is zero-shot learning?
Zero-shot learning is a model’s ability to handle a task or category with no training examples of it, working only from a description. A language model asked to sort emails into categories defined in one sentence each is working zero-shot. It relies on general knowledge from pre-training and on the link between the description and what it has seen before.
Large language models made zero-shot learning an everyday tool. Writing “Classify this ticket as hardware, software, access, or billing” in a prompt is zero-shot classification. When a few labeled examples are added to the prompt, it becomes few-shot learning, and when hundreds are used to retrain the model, it becomes fine-tuning.
Zero-shot learning in practice
A managed service provider wants incoming helpdesk tickets routed to the right queue. The team writes a prompt that names the eight queues and describes what belongs in each, and a language model routes tickets from the first day. The team logs the model’s choices, has dispatchers correct the wrong ones, and reviews the results after a month to decide whether few-shot examples or a fine-tuned model are worth the effort.
What to watch for
- Zero-shot accuracy depends heavily on how the categories are described, so test several phrasings.
- Categories specific to the business, such as internal product codes, are where zero-shot fails first.
- Measure against a labeled sample anyway, because a model that sounds confident can still be wrong.