Few-Shot Prompting

Few-shot prompting places several input–output examples in the context so the model learns the task in-context, without any weight updates. Brown et al. (2020) showed this ability emerges with scale, and that adding a few examples often closes much of the gap to fine-tuned systems for common tasks.

Examples are most useful when the desired format is specific or unusual — exactly the cases where zero-shot tends to wobble.