In-Context Learning is a capability of large language models (LLMs) that allows them to learn new tasks or behaviors at inference time, without any parameter updates or fine-tuning. Instead of retraining the model, users provide task instructions, examples, or demonstrations directly within the input prompt, and the model uses that context to generate appropriate responses.

For example, given this prompt:

Translate to French:

‍
Dog → Chien
Cat → Chat
Bird → ...

The model will likely complete with “Oiseau” — not because it was specifically trained on this prompt, but because it can generalize patterns from the examples provided in the context.

In-context learning supports:

• Zero-shot learning: The model performs a task based on a prompt describing the task (no examples needed).
• Few-shot learning: The model is given a handful of examples to guide its behavior.
• Instruction following: The model learns to follow explicit task instructions embedded in the prompt.

This behavior emerges in large transformer-based models as their scale increases, and is a defining characteristic of modern LLMs like GPT-3 and GPT-4. It enables rapid adaptation to new tasks without retraining, making LLMs flexible and powerful for a wide variety of real-world applications.