in-context learning (ICL)

Gold definitionUpdated Apr 2, 2026

Definition

In-context learning (ICL) is a capability of large language models (LLMs) to learn new tasks or adapt their behavior by observing examples provided directly within the input prompt, without requiring explicit weight updates or fine-tuning.

At a glance

Executive summary

In-context learning allows large AI models to learn new tasks or adjust their behavior simply by being shown examples directly within the text prompt, without needing to be retrained. This makes them incredibly flexible and able to handle many different jobs on the fly, saving time and computing power.

TL;DR

Large AI models can learn new tasks from examples given in the prompt itself, making them adaptable without needing to be retrained.

Key points

LLMs adapt to new tasks by processing examples within the input prompt, without weight updates.
Solves the problem of needing to fine-tune models for every new task, enabling rapid adaptation.
Used by researchers and ML engineers in NLP applications requiring flexible model behavior.
Differs from fine-tuning by not altering model weights, relying solely on pre-trained knowledge and prompt context.
Ongoing research focuses on understanding its underlying mechanisms and improving prompt engineering techniques.

Use cases

Few-shot text classification: Classifying sentiment or topic with only a few labeled examples in the prompt.
Machine translation: Translating sentences into a specific style or dialect by providing examples of desired translations.
Code generation: Generating code snippets in a particular programming language or framework by showing example code.
Question Answering: Answering questions based on specific document sections provided in the prompt, mimicking a retrieval-augmented generation setup.
Creative writing: Guiding an LLM to write in a specific author's style by including examples of their work in the prompt.

Also known as

ICL, few-shot learning (in LLMs), prompt-based learning, demonstration learning