In-context learning can adapt a model's output during inference without explicit gradient updates or parameter modification.

The retrieved theoretical literature defines in-context learning as adaptation driven by examples or information placed in the prompt. The adaptation occurs during inference, while the model's stored parameters remain unchanged. This separates behavioural updating from ordinary training: the model alters its computation in response to context, not its persistent weights. The distinction is operationally important because improved task performance alone does not establish that stable learning has occurred.