In-Context Learning

In-context learning is the ability of a large language model to perform a new task from examples given in the prompt alone — no fine-tuning, no gradient updates. The “learning” happens entirely within the forward pass over the context.

Worked example: show the model three examples of “English → French” in the prompt and it will translate a fourth, having inferred the pattern from context; this is what few-shot prompting exploits. Gotcha: it is bounded by the context window and can be brittle — sensitive to example order and formatting — and it does not persist: the model has not truly learned the task, it just conditioned on your examples for this one call.