Overview
Zero-shot learning relies on the model's ability to generalize from its existing knowledge. For example, a model that knows what a 'horse' and 'stripes' are might be able to identify a 'zebra' without ever having seen one.
How it Works
Models often use semantic descriptions or attribute-based representations to bridge the gap between seen and unseen classes.
Importance in LLMs
Modern LLMs are excellent at zero-shot tasks, such as translating a language they weren't specifically fine-tuned for, simply by following instructions.