Few-Shot Learning
Theory
Few-shot prompting places worked examples directly in the prompt context. The model pattern-matches them and is likely to continue that pattern — no weight updates, no retraining.
Zero-shot — instruction only
Prompt: "Classify this review."
Output: Sentiment: 😊 (informal — guess) — format invented per call.
Few-shot — 3 demos
Demos: "Fast shipping" → {label: positive}
Demos: "Broken on arrival" → {label: negative}
...
Output: {label: positive, confidence: high} — locked shape.
Zero-shot0 demos · obvious tasks
One-shot1 demo · light format cue
Few-shot2–8 demos · unusual shapes
Coming next: chain-of-thought adds a reasoning trace to each example, which tends to lift accuracy on multi-step tasks where the final answer alone leaves intermediate steps ambiguous.