How is in-context learning different from fine-tuning?

In-context learning happens at inference time through the prompt - no model changes. Fine-tuning modifies the model's weights through additional training. ICL is faster and cheaper; fine-tuning produces more persistent changes.

Why do larger models learn better in-context?

In-context learning emerges from scale - larger models develop better pattern recognition and generalisation. Smaller models may need fine-tuning to achieve similar task-specific performance.

Is in-context learning permanent?

No, it only lasts for the current request. The model doesn't retain information across requests. For persistent learning, you need fine-tuning, or to include the context in each request.

What are the limits of in-context learning?

Limits include: context window size (can't fit unlimited examples), tasks requiring deep domain expertise, tasks needing structural changes to model behaviour, and very novel tasks far from training distribution.

Book Free Assessment Calculator

In-Context Learning

The ability of LLMs to learn and adapt their behaviour based on examples and instructions provided in the prompt, without model updates.

In-Depth Explanation

In-context learning (ICL) is the remarkable ability of large language models to adapt to new tasks based solely on examples provided in the prompt - no gradient updates or model changes required.

How in-context learning works:

The model receives a prompt with instructions and/or examples
Through its attention mechanism, it identifies patterns in the examples
It applies these patterns to generate appropriate outputs
All of this happens at inference time, using the model's existing weights

Forms of in-context learning:

Zero-shot: Task description only, no examples
One-shot: Single example provided
Few-shot: Multiple examples demonstrating the pattern
Instruction following: Detailed natural language instructions

Why ICL is powerful:

Instant adaptation without training
Easy to iterate and experiment
Works across diverse domains
Accessible without ML expertise
Preserves base model capabilities

Research shows ICL emerges with scale - smaller models don't exhibit this ability as strongly.

Business Context

In-context learning enables rapid customisation - you can change AI behaviour instantly by updating prompts rather than retraining models.

How Clever Ops Uses This

We leverage in-context learning extensively at Clever Ops, enabling rapid prototyping and iteration for Australian business clients without the delays of model training.

Example Use Case

"Providing 3 examples of your preferred email format so the AI mimics your style for all future emails."

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

Implementation Partner Incident Management