Few-shot / Zero-shot Learning
Updated 11.08.20261 min
Few-shot means putting a handful of example answers into the request so the model copies their format. Zero-shot is the same request with instructions only and no examples.
How it works
The model learns nothing in the process; the weights do not change. The examples simply sit in the context, and the continuation naturally inherits their structure. Strictly it is not learning but a very precise statement of the format.
Which is why a couple of samples beat a paragraph of explanation. Showing what an answer should look like is almost always shorter and more reliable than describing it.
The flip side: the model copies what you did not intend as well. If every example starts with a verb or fits two lines, it will treat that as part of the brief.
How to choose examples
- Two or three is enoughMore examples consume context and rarely improve the result past the third.
- Include a hard caseOne example containing an exception conveys the task's boundaries better than three easy ones.
- Watch for accidental patternsMatching length, order or tone across all examples will be reproduced as a requirement.
- If you need many, that is fine-tuningDozens of examples in every request cost money and space. Persistent behaviour is cheaper to bake in by fine-tuning.