Few-shot learning provides examples in the prompt to guide LLM behavior. The model learns the pattern from examples and applies it to new inputs without any training.
Types of In-Context Learning
- Zero-shot: No examples, just instructions
- One-shot: Single example
- Few-shot: 2-5 examples typically
- Many-shot: More examples (uses more tokens)
Best Practices
- Use diverse, representative examples
- Match example format to desired output
- Order examples thoughtfully
- Balance token cost vs accuracy
How many examples do I need?
Usually 2-5 examples work well. More examples improve consistency but cost more tokens. Test to find the right balance.
Practical example
A classifier receives a few examples showing how ambiguous support requests should be labeled.
What to check
Keep evaluation cases separate from prompt examples and test changes to example order and wording.
Limits and tradeoffs
Examples consume context and can introduce bias; in-context learning does not necessarily update model weights.
Maintained by DriftRail. AI assisted the October 2026 content and template revision. Examples are illustrative; they are not measured product benchmarks. Suggest a correction.