Zero-shot learning
An LLM's ability to perform a task it has never been explicitly trained on or shown examples of — relying entirely on the model's pre-training and the prompt instructions.
Zero-shot is the test of whether an LLM has internalized a capability. You describe the task in natural language; the model performs it without any examples. "Classify this email as spam or not" with no examples shown — that is zero-shot.
Frontier models in 2026 handle most common tasks zero-shot at production quality. Classification, extraction, summarization, translation, basic reasoning — all work zero-shot with a well-written prompt. Few-shot still helps on edge cases and unusual formats.
The practical lesson: try zero-shot first. If accuracy is acceptable, ship. Add few-shot examples only when zero-shot misses cases, and fine-tune only when few-shot plateaus. Most teams over-engineer prompts that work fine without examples.
Frequently asked
When does zero-shot fail?+
On tasks with idiosyncratic format requirements, on highly domain-specific terminology, and on tasks the model has not seen variants of during training. Add few-shot examples or fine-tune in those cases.
Are reasoning models always zero-shot?+
They are designed to work zero-shot on hard problems, and they generally do. But adding examples still helps on tasks where format consistency matters more than reasoning depth.