Zero-shot learning, as the term is used in the LLM context, refers to a large language model's ability to follow an instruction and perform a task correctly without being shown any worked examples within the prompt. The model relies entirely on knowledge and reasoning patterns absorbed during pretraining rather than in-context demonstrations. Strong zero-shot capability is widely regarded as a benchmark for how generally capable a model is, because it demands broad world knowledge and robust instruction-following.

Think of it like handing someone a job description and asking them to start immediately, with no onboarding examples — a highly experienced hire succeeds; a narrow specialist struggles. The richer the pretraining, the more tasks the model can handle cold.

A canonical zero-shot scenario: you present an LLM with a classification schema using only a natural-language prompt:

Label each sentence as one of: Regulatory, Technical, or Financial.
Sentence: "The interest rate swap matures in Q3."

No labelled sentences precede it. State-of-the-art models handle this reliably; smaller models often misclassify or ignore the schema entirely.

Zero-shot learning is the baseline case within the broader spectrum of prompt engineering, sitting below few-shot prompting and chain of thought in terms of guidance provided to the model. It is enabled by the same transformer attention mechanisms that give LLMs their contextual flexibility, and its limits are part of what motivates RAG and fine-tuning when domain-specific accuracy matters.

You will encounter zero-shot learning benchmarks in model evaluation papers — datasets such as BIG-Bench and MMLU report zero-shot scores — and in everyday use whenever you ask an LLM to classify, translate, or summarise without giving it examples first.