When should I use zero-shot vs few-shot?

Start with zero-shot for simple, well-defined tasks. Add examples (few-shot) if results are inconsistent, you need specific formatting, or the task is specialised. Few-shot costs more tokens but usually improves accuracy.

Why does zero-shot work?

LLMs learn broad patterns during pre-training that transfer to new tasks. When you describe a task in natural language, the model can often figure out what's needed based on similar patterns it learned during training.

Do larger models do better zero-shot?

Generally yes - zero-shot capabilities scale with model size. Larger models have seen more training data and developed better generalisation. GPT-4 and Claude 3 Opus are significantly better at zero-shot than smaller models.

What if zero-shot gives inconsistent results?

Try: clearer task descriptions, specific output format instructions, lower temperature for consistency, or add 1-3 examples (few-shot). Inconsistency often means the task isn't fully defined from the model's perspective.

Book Free Assessment Calculator

Zero-Shot Learning

Performing tasks without any examples in the prompt, relying solely on the model's pre-trained knowledge and the task description.

In-Depth Explanation

Zero-shot learning is the ability of AI models to perform tasks they haven't been explicitly shown examples of, using only a natural language description of what's needed. This demonstrates the generalisation power of modern LLMs.

How zero-shot works:

Provide a clear task description
Model draws on pre-trained knowledge
Generates appropriate output without examples
Works because training exposed the model to similar patterns

Zero-shot vs few-shot:

Zero-shot: Task description only, no examples
One-shot: Single example provided
Few-shot: Multiple examples provided

When zero-shot works well:

Common, well-defined tasks
Clear, unambiguous instructions
Tasks similar to training distribution
Larger, more capable models

When to add examples (few-shot):

Complex or nuanced tasks
Specific formatting requirements
Domain-specific applications
When zero-shot produces inconsistent results

The emergence of zero-shot capability is one of the remarkable properties of large-scale language models - they can generalise to new tasks without explicit training on them.

Business Context

Zero-shot works for straightforward tasks. For complex or specialised tasks, few-shot learning typically performs better.

How Clever Ops Uses This

We start with zero-shot approaches for simplicity. When more precision is needed, we add examples strategically. This iterative approach helps us find the optimal balance for each use case.

Example Use Case

"Asking "Classify this text as positive, negative, or neutral" without any examples - the model understands the task from the description alone."

Frequently Asked Questions

Learn More

Prompt Engineering Best Practices: Master the Art of AI Communication

Learn proven techniques for writing effective prompts that consistently produce high-quality results from LLMs. Includes practical examples, templates, and testing strategies for production applications.

Read article

Zero Trust Architecture