What is zero-shot

What is zero-shot refers to the machine learning capability where AI systems perform tasks or classify categories they have never encountered during training, leveraging learned representations and semantic knowledge to generalize beyond their training distribution. This paradigm enables models to handle novel scenarios without requiring task-specific examples by exploiting semantic embeddings, attribute-based learning, and cross-modal knowledge transfer. Zero-shot approaches utilize techniques like mapping textual descriptions to visual features, leveraging pre-trained embeddings, and exploiting compositional understanding of concepts. Common implementations include vision-language models classifying unseen object categories, language models following novel instructions, and recommendation systems handling new items. The capability emerges from models learning generalizable patterns and relationships that transfer across domains. For AI agents, zero-shot capabilities provide immediate deployment to new tasks and cost-effective scaling.