In-context learning

When a model does a new task from instructions or examples written in its input, without any change to its weights. It works only during that one input; nothing is stored afterwards.