Few-Shot Learning
What is it?
Few-shot learning is giving a model a handful of examples in the prompt so it can pick up the pattern and handle a new task without any retraining.
Explain like I'm 5
Why was it created?
Retraining a model for every small task is expensive. Large language models turned out to learn tasks from just a few in-prompt examples, so few-shot prompting became a fast, free alternative.
Where is it used?
- Classification and labeling
- Formatting output consistently
- Style or tone matching
- Data extraction
Why should developers care?
It's one of the quickest ways to steer a model toward the exact format and behavior you want, without fine-tuning or extra infrastructure.
How does it work?
You include several input-output examples in the prompt, then the new input. The model infers the pattern from the examples (in-context learning) and applies it to the new input — no weights change. Zero-shot uses none, one-shot uses one, few-shot uses several.
Real-world example
To classify support tickets, you show three tickets each labeled 'billing' or 'bug', then a new ticket — the model labels it in the same scheme.
Common use cases
- Consistent output format
- Quick classification
- Teaching a niche pattern
- Bootstrapping before fine-tuning
Advantages
- No training or infrastructure needed
- Fast to iterate
- Strongly steers format and behavior
- Works with any capable model
Disadvantages
- Examples consume context window
- Sensitive to example choice and order
- Less reliable than fine-tuning at scale
- Adds cost per call
When should you use it?
When you need to steer behavior or format quickly and a few examples capture the pattern.
When should you avoid it?
When you have many examples and high volume (fine-tuning is cheaper per call) or when examples won't fit the context window.
Alternatives
Related terms
Interview questions
Beginner
- What is the difference between zero-shot and few-shot?
- Where do the examples go in few-shot learning?
Intermediate
- Why can example order affect results?
- When is fine-tuning better than few-shot?
Senior
- How does in-context learning differ from gradient-based learning?
- How would you select examples for a few-shot prompt at scale?
Common misconceptions
- "Few-shot updates the model" — nothing is trained; it's all in the prompt.
- "More examples always help" — they cost context and can even hurt if poorly chosen.
Fun facts
- GPT-3's paper was literally titled 'Language Models are Few-Shot Learners'.
- The examples are sometimes called 'shots'.
Timeline
- 2020 — GPT-3 demonstrates strong few-shot in-context learning
- 2020s — Few-shot prompting becomes everyday practice
Learning resources
Quick summary
Few-shot learning puts a few examples in the prompt so a model infers the pattern and handles a new task with no retraining.
Cheat sheet
- Examples in the prompt, no training
- Zero/one/few = 0/1/several examples
- Great for format and classification
- Costs context; order matters