Despite the surprising few-shot performance of in-context learning (ICL), it
is still a common practice to randomly sample examples to serve as context.
This paper advocates a new principle for ICL: self-adaptive in-context
learning. The self-adaption mechanism is introduced to help ea