Zero-shot Learning

The ability of an AI model to complete a task without having received any specific training examples for that task.

How it works

Zero-shot capability emerges from the breadth of pre-training. A model trained on vast, diverse text implicitly learns the structure of many tasks: instruction following, question answering, translation, summarisation, and more. At inference, a clear instruction in the prompt — without any examples — is sufficient to elicit the correct behaviour. The model's generalisation comes from having seen related patterns across millions of documents, not from task-specific training examples.

Why it matters

Zero-shot performance is the most practical measure of a model's general intelligence. It determines how useful the model is 'out of the box' for novel tasks — without any dataset collection or fine-tuning investment. Strong zero-shot performance is what makes LLMs so broadly applicable across industries and domains. Improvements in zero-shot capability — driven by scale, better pre-training data, and instruction tuning — are the primary driver of the rapid expansion of LLM use cases.