Zero-Shot Learning

Zero-shot learning means asking a model to do something it has never been explicitly shown how to do, relying entirely on its pre-trained knowledge. For example: "Classify this email as spam or not spam: [email text]". No examples needed; the model understands the task from the instruction alone. Modern LLMs are remarkably good at zero-shot tasks. GPT-5.2, Claude Opus 4.6, and Gemini 3.0 Pro can handle most straightforward tasks without examples. Zero-shot is the fastest way to prototype: write a clear instruction and test it. If accuracy is insufficient, add few-shot examples. In practice, zero-shot works well for: general knowledge tasks, summarization, translation, simple classification, and creative generation. It struggles with: domain-specific formatting, nuanced categorization, and tasks where the output format is non-obvious. Start zero-shot, measure accuracy, and escalate to few-shot or fine-tuning only when needed.