Putting examples in the prompt instead of training on them.
这两项技术尚未记录可对比的属性。
由上述差异按固定规则生成，不是模型生成的。
Choose Few-Shot Learning if…
已记录的属性中，没有任何一项显示它更有优势。
Choose Zero-Shot if…
已记录的属性中，没有任何一项显示它更有优势。
页面保持可访问，但在积累足够的带来源细节之前不纳入搜索索引（PRD SEO-08）。
technique · Prompting
Few-shot learning supplies a handful of worked examples in the prompt so the model infers the task from them without any weight update. It solves the problem that fine-tuning needs labelled data, a training run and a deployment per task, while many tasks are simple enough to demonstrate in a paragraph. It exists because it emerged from scale rather than design: the capability appeared in large models that were never trained for it.
concept · Prompting
Zero-shot means asking a model to do something it was not trained on and giving it no examples - only a description of the task. It solves the problem that every supervised approach needs labelled data for each new task, which does not scale to tasks nobody has collected data for. It exists as the capability that made general-purpose models useful, and as the upper bound that few-shot prompting is measured against.
