What is zero-shot prompting? Defining tasks without demonstrations
Understand zero-shot prompting through email classification, distinguishing no demonstrations from no training, with labels, uncertainty, evaluation, and when to add examples.
Understand the concept
Zero-shot prompting describes a task without supplying demonstrations of its answers in the current request. It does not mean an untrained model or prove absence of similar training data. Context and factual references can still be included. Ambiguous tasks still require clear output criteria and evaluation on new examples.
Define labels rather than asking only for classification
For quotation inquiries versus support issues, define pre-purchase pricing and post-purchase use problems before supplying an email. Without demonstrated answers this remains zero-shot. Bare label names can yield inconsistent decisions on similar messages.
Specify multiple labels or a primary intent for mixed emails, plus a policy for missing information. Otherwise the model may force a label to fit the format. This is evaluation design without real email access or measured accuracy.
Zero-shot does not establish absence of related training
The GPT-3 study distinguishes settings by demonstrations in the prompt, not proof that training lacked similar tasks. With undisclosed training data, prior exposure remains unknown. Zero-shot results should not be described as absolute learning from nothing.
Supplying policy text does not necessarily make prompting few-shot: references differ from demonstrated input-output pairs. Distinguish missing policy knowledge from missing output-format guidance when diagnosing errors.
Use unseen cases to decide whether demonstrations help
Evaluate labeled emails for boundary and format errors. If zero-shot meets the need, added examples may not justify capacity. For persistent ambiguity, add a targeted demonstration and compare on separate inputs rather than only the new example.
Demonstrations or not, models can invent details. Check original evidence for factual tasks and error costs for classification. This explains prompt settings without declaring one universally superior or treating prompting style as proof of general capability.
Checked sources
These sources support the concepts and methods described here. Examples and test scope are labeled separately; check the version of the tools you use.