Definition
Zero-shot means the model gets a task description but no worked examples. Modern LLMs often succeed via pretraining knowledge and instruction tuning.
In classical ML, zero-shot can mean classifying unseen classes using semantic descriptions.
In simple terms
Being asked to assemble furniture with only the written instructions, no demo video, and still succeeding.
Where you see it
- "Summarize this article in Somali" with no sample summaries.
- Zero-shot image classification with CLIP-style models.
How it works
1.Describe the task
Clear instructions and constraints.
2.Provide the input
Text, image, or other modality.
3.Model generalizes
Uses pretrained skills to comply.
Why it matters
- Zero-shot capability is why general-purpose models can jump into new workflows instantly.
Often confused
Zero-shot means the model never saw related data in training.
It means no examples in this request, pretraining may still include similar tasks.