Few-Shot Learning
Few-shot learning is giving a model a handful of worked examples in the prompt so it picks up the pattern and applies it to your task — no retraining. Zero-shot means no examples (just the instruction); few-shot adds a few; both work because large models learn from context at inference time.
Also known as: few-shot prompting, zero-shot learning, in-context learning
You can steer a model two ways without touching its weights. Zero-shot gives it only the instruction (“classify this review’s sentiment”). Few-shot adds a few labeled examples first, so the model infers the pattern — the format, the edge cases, the style — and matches it. This works because large models do in-context learning: they adapt to examples in the prompt at inference time, no fine-tuning required.
It’s the cheapest customization lever you have, and it pairs directly with chain-of-thought prompting — few-shot examples that show the reasoning teach the model to reason, not just to answer. The trade-off is that examples eat context window and tokens, and as models improve, zero-shot often closes the gap. The practical rule: try zero-shot first, add examples when the model needs to see the pattern, and reach for fine-tuning only when no number of examples gets you there.