Jargonautica
Controlling and Extending Behaviour at Run Time

Zero-shot, One-shot and Few-shot

Also called: Zero-shot, Few-shot, One-shot, In-context Learning

How many examples you give the model before asking it to do something. Zero-shot means you simply describe the task. Few-shot means you show it two or three worked examples first. Showing examples is often the fastest way to improve results, especially when you want a specific format, and it usually beats trying to describe that format in words.

Technical view

This is in-context learning: the model adapts its behaviour from examples in the prompt without any parameter updates. It was one of the notable emergent properties of scale. Example selection and ordering both affect results measurably, and gains typically plateau after a handful of examples. Instruction-tuned models have narrowed the gap between zero-shot and few-shot performance considerably compared with early base models.

Part of AI Terminology Index, section: Controlling and Extending Behaviour at Run Time.

How it connects

Zero-shot, One-shot and Few-shot links to 8 other terms. The faint lines are links between those terms. See it in the full graph.

Fine-tuningFine-tuningChain of ThoughtChain of ThoughtEmergent CapabilitiesEmergent Capabiliti…Prompt EngineeringPrompt EngineeringInstruction TuningInstruction TuningInduction HeadInduction HeadContext WindowContext WindowTime Series Foundation ModelTime Series Foundat…Zero-shot, One-shot and Few-shot