Zero-Shot Learning

Zero-shot learning means asking a model to do something using only an instruction, with no examples at all. It relies entirely on what the model already learned.

That this works is the surprising part. A model never trained to classify your specific categories can often do it from a clear description, because the underlying concepts were in its training data.

Think of it like this. Think of asking a well-read colleague to sort documents into categories they have never sorted before. No training needed, because they already understand what the categories mean.

How it works

• Instruction only. Describe the task in plain language.
• The model draws on pretraining, where it saw enough related material to generalise.
• Clarity carries all the weight. With no examples to disambiguate, the wording is everything.
• Cheapest option, since no example tokens are sent.

When to use it

• Common, well-understood tasks: summarise, translate, classify sentiment.
• When output format is simple or does not matter much.
• When token cost matters, since it is the cheapest possible prompt.
• Try it first. If zero-shot works, few-shot is wasted money.

Common mistakes

• "Zero-shot means the model is guessing." It is applying knowledge from pretraining. On familiar tasks it is often as good as few-shot.
• "It should work for anything." Unusual formats and domain-specific conventions need examples. That is exactly what few-shot is for.
• "A vague instruction is fine." With nothing to imitate, ambiguity in your wording goes straight into the output.