Zero-Shot Prompting

Zero-shot prompting is the practice of asking a language model to perform a task by describing it directly in natural language, without providing any worked examples of correct input-output pairs. It’s the simplest form of prompting — just an instruction — and stands in contrast to few-shot prompting, which includes several examples to demonstrate the desired pattern before the actual task input.

Why It Works

Modern instruction-tuned LLMs are explicitly trained (via supervised fine-tuning and reinforcement learning from human feedback) to follow natural-language instructions directly, which makes zero-shot prompting far more reliable today than it was with earlier, purely next-token-predicting base models. A well-instructed frontier model can often perform reasonably well on a new task — summarization, classification, translation, extraction — from a clear zero-shot instruction alone, without needing any examples.

When Zero-Shot Falls Short

Zero-shot prompting tends to struggle when the desired output format is unusual, the task requires following a subtle stylistic convention, or the boundary between correct and incorrect outputs is ambiguous without a concrete example. In these cases, few-shot prompting — showing 2-5 examples of the exact input-output pattern desired — typically produces more consistent, correctly formatted results, at the cost of using more tokens per request.

Practical Guidance

A common workflow is to start with a zero-shot prompt for simplicity and speed, and add few-shot examples only if evaluation shows the zero-shot version producing inconsistent formatting or missing nuances in the task definition — since each added example increases token cost and latency on every subsequent request using that prompt.

Zero-Shot Chain-of-Thought

An important variant worth distinguishing is zero-shot chain-of-thought prompting, which combines a zero-shot instruction with a simple nudge like “let’s think step by step” — this remains zero-shot in the sense that no worked examples are provided, but it still meaningfully improves reasoning accuracy on many tasks compared to a bare zero-shot instruction asking directly for a final answer. This distinction matters because it shows zero-shot and reasoning-elicitation techniques aren’t mutually exclusive: a prompt can be zero-shot (no examples) while still using structural techniques to improve the quality of the model’s response, rather than “zero-shot” simply meaning “the simplest possible prompt.”

This is a useful distinction to keep in mind when reading benchmark papers, which often report separate zero-shot and few-shot scores for the same model, since the gap between them reveals how much a given task benefits from demonstrated examples versus how well the model already generalizes from instructions alone.