Review Zero-Shot Learning

Zero-shot means asking a model to do a task from instructions alone, with no examples. Few-shot means including a small number of worked examples in the same request. In neither case is the model retrained — the examples only shape the answer to that one request.

In a zero-shot request, the model gets the task but no demonstration. In a few-shot request, the prompt includes several examples showing the desired input-output pattern.

Examples can improve consistency quickly, but they consume context and may introduce unwanted patterns. Teams still need representative tests and human review.

A prompt asks a model to label feedback by category and includes three reviewed examples before presenting a new comment.

Few-shot learning in a prompt does not permanently retrain the model. The examples guide the current context unless a separate learning system stores or trains on them.

Zero-shot tells; few-shot shows.

* [[prompt|Prompt]] * [[prompt-engineering|Prompt Engineering]] * [[context-window|Context Window]] * [[fine-tuning|Fine-Tuning]]

* [Brown et al., *Language Models are Few-Shot Learners*](https://arxiv.org/abs/2005.14165) — primary paper evaluating zero-shot, one-shot, and few-shot task performance from instructions and in-context demonstrations without gradient updates.

* Contrast temporary in-context examples with model changes from [[fine-tuning|Fine-Tuning]].