Few-Shot Prompting

Sources:

• Prompt Engineering Guide - Few-Shot

Overview

While LLMs demonstrate remarkable Zero-Shot Prompting capabilities, they fall short on complex tasks. Few-shot prompting enables in-context learning by providing demonstrations in the prompt to steer the model toward better performance.

The demonstrations serve as conditioning for subsequent examples where the model generates a response.

Example

From Brown et al. 2020:

Prompt:

A "whatpu" is a small, furry animal native to Tanzania. An example of a sentence that uses the word whatpu is:
We were traveling in Africa and we saw these very cute whatpus.

To do a "farduddle" means to jump up and down really fast. An example of a sentence that uses the word farduddle is:

Output:

When we won the game, we all started to farduddle in celebration.

When to Use Few-Shot

Scenario	Approach
Simple, common tasks	Zero-shot first
Model struggles with zero-shot	Add 1-3 examples
Complex reasoning or formatting	3-5+ examples
Domain-specific knowledge	Include domain examples

Best Practices

1. Diverse examples: Cover different cases/edge scenarios
2. Consistent format: Use identical structure across examples
3. Quality over quantity: Well-crafted examples beat many poor ones
4. Order matters: Place similar examples near the query
5. Label balance: Include examples of different output types

Limitations

• Token limit constrains number of examples
• Examples can introduce bias if not diverse
• May not generalize to very different inputs

Related Techniques

Technique	Description
Zero-Shot Prompting	No examples provided
Chain-of-Thought Prompting	Include reasoning steps
Self-Consistency	Sample multiple reasoning paths

---

See Also

• Zero-Shot Prompting
• Chain-of-Thought Prompting
• MOC - Artificial Intelligence

---

(c) No Clocks, LLC | 2024