Zero-shot learning
Zero-shot learning is an LLM's ability to perform a task correctly without a single example in the prompt — entirely from pre-training.
In a zero-shot prompt you ask the model to perform the task directly, with no examples. Modern LLMs (GPT-4, Claude 3.7) perform surprisingly well zero-shot on many tasks because they've seen millions of similar patterns during training. Efficient for simple, unambiguous tasks; for nuance or company-specific output, few-shot or fine-tuning is needed.
Example
Zero-shot: 'Classify this review as positive or negative: [review]'. No examples, no explanation — the model knows what sentiment classification is and does it immediately.
Frequently asked questions
Zero-shot or few-shot?
Zero-shot for standard tasks (summarise, translate, classify). Few-shot once you need a specific format, tone or domain context.
When does zero-shot fail?
Niche jargon, company-specific formats, rare languages, or when the model must deviate from its default. Then: few-shot or fine-tune.
Related terms
Further reading
- → Our service: GEO