In-Context Learning (ICL)

A model's ability to adapt behavior to instructions, demonstrations, or patterns present in the current call context without updating its weights. The effect lasts while that context is available and can guide classification, format, language, or temporary mappings.

ICL is not persistent learning. Example order, quality, and distribution change the outcome, and a poor demonstration can induce the wrong rule. Comparison with zero-shot prompting and fine-tuning should consider accuracy, context cost, maintenance, and stability beyond the demonstrated cases.