Zero-Shot Learning

Zero-shot learning is a model's ability to perform a task it was not specifically trained for and has seen no examples of, working only from a description of what to do. With a large language model, a zero-shot prompt states the task ('classify these reviews as positive, negative or mixed') and supplies the input, without any worked examples.

The term has two related uses. In classic machine learning it means recognizing classes that never appeared in training by relying on descriptions or attributes, for example an image model labeling a photo as 'a folding e-bike' by comparing it with text descriptions. In LLM prompting it is one end of a spectrum defined in OpenAI's GPT-3 paper, Language Models are Few-Shot Learners (2020): zero-shot (instruction only), one-shot (one example) and few-shot (a handful of examples in the prompt). In that paper, GPT-3 with 175 billion parameters did markedly better with a few examples than without any, which is why few-shot prompting became the default for a while.

Two developments made zero-shot prompting the normal way to use LLMs. Instruction tuning and RLHF trained models to follow plain task descriptions. And researchers found that reasoning could be triggered without examples: Kojima et al. (2022) showed that adding 'Let's think step by step' raised InstructGPT's accuracy on the MultiArith benchmark from 17.7% to 78.7% and on GSM8K from 10.4% to 40.7%, with no worked examples at all. Today's reasoning models do that chain-of-thought step internally, so frontier models such as GPT-6, Claude Opus 5.5 and Gemini 4 Argon are usually prompted zero-shot for everyday tasks.

Zero-shot has limits. It struggles when the output format is unusual, when labels have a company-specific meaning ('qualified lead' means something different in every CRM), or when the instruction is ambiguous. In those cases a few examples or fine-tuning still help. It also adds no knowledge: a model working zero-shot on a topic past its knowledge cutoff can only fill the gap by retrieving information or by guessing, and guessing is where hallucinations come from.

For AI search, nearly every user prompt is a zero-shot task. Someone asks a new question with no examples, and the model decides on its own how to break it into searches (query fan-out), which sources to trust and how to structure the answer. For long-tail and novel prompts, nobody has optimized a page for that exact wording, so the model relies on whatever passages define the concepts and state the relationships it needs. Content that names its subject, defines its terms and makes comparisons explicit gives a model usable material for questions no one anticipated. The same idea applies to the measurement side: AI visibility tools often use zero-shot LLM classification to label brand mentions, sentiment or prompt intent, so it is worth checking a hand-labeled sample against those labels before relying on them. See prompt engineering for how instructions and examples are combined in practice.