What Is In-Context Learning?

In-context learning is the ability of a large language model to adapt its behavior for a new task based solely on examples or instructions provided within its input prompt, without updating any of its internal weights. The model is not “learning” in the traditional sense of backpropagation adjusting parameters; instead, it is using patterns already encoded from its original training to recognize and follow the structure of the examples it is shown right there in the conversation.

This stands in sharp contrast to how machine learning models traditionally improved at a task, which required collecting a dataset and retraining or fine-tuning the model through many iterations. In-context learning happens entirely at inference time, meaning while the model is actively generating a response, and it disappears the moment the conversation ends, since no permanent change is made to the model itself.

How In-Context Learning Works

In-context learning typically shows up in a few recognizable patterns, often described by how many examples are provided:

Zero-shot: The model is simply given an instruction, with no examples at all, and must rely entirely on patterns from its original training to produce a reasonable response.

One-shot: The model is given a single example of the task before being asked to perform it on new input, giving it just one point of reference for the expected format or style.

Few-shot: The model is given several examples, typically between two and a few dozen, demonstrating the task before being asked to continue the pattern. More examples generally give the model a clearer sense of the exact structure or reasoning expected.

What makes this possible is the underlying attention mechanism inside transformer-based models, which lets the model weigh and relate every part of its input, including the examples provided, to every other part when generating its response. The examples in the prompt effectively act as a temporary, situational guide that the attention mechanism can draw on for that single generation, even though none of the model’s underlying weights change as a result.

A Simple Analogy

Imagine handing a skilled translator a short style guide right before they start a new project, showing a few sentences translated the way you prefer, perhaps more formal, or using a specific regional phrase. The translator does not need to go back to language school or relearn grammar; they already have all the underlying skill. They simply adjust their approach for this one project based on the examples you just showed them, then return to their normal style for the next assignment. That immediate, temporary adjustment based only on a few shown examples is exactly what in-context learning looks like in a language model.

In-Context Learning vs. Fine-Tuning

It helps to be clear about how in-context learning differs from other ways models are adapted to new tasks, particularly transfer learning and fine-tuning. Fine-tuning permanently updates a model’s weights using a training dataset, through repeated passes of backpropagation guided by a loss function, and those changes persist for every future use of that model. In-context learning changes nothing permanent at all; it is a temporary behavioral adjustment that exists only for the duration of a single prompt and response.

This makes in-context learning remarkably fast and flexible. There is no need to collect a large labeled dataset, no training run, and no risk of overfitting to a small set of examples in any lasting way. On the other hand, it also has real limits: a model can only hold so much information in its context window at once, and the adaptation vanishes between conversations, meaning the same guidance has to be re-supplied every time it is needed.

Why In-Context Learning Matters

In-context learning is a big part of why large language models feel so versatile in everyday use. A single model can be asked to write in a particular tone, follow a specific output format, perform a basic classification task, or even imitate a style of reasoning, all just by including a handful of well-chosen examples in the prompt, without anyone needing to train a specialized model for each situation.

This capability emerged somewhat unexpectedly as language models grew larger and were trained on increasingly diverse data; it was not explicitly designed into early models but rather appeared as a byproduct of scale. Researchers continue to study why and how it works so reliably, since the model is effectively solving a new task on the fly using nothing but its existing, general-purpose knowledge and the brief hints given in a prompt. Understanding in-context learning helps explain why writing a clearer prompt, with better examples, so often produces noticeably better results, even when nothing about the underlying model itself has changed.