Finding Golden Examples: A Smarter Approach to In-Context Learning

In-Context Learning (ICL) improves Large Language Model performance, but its effectiveness depends heavily on the quality of the examples provided. A method from a Google DeepMind paper, called AuPair, offers a systematic approach to finding the best examples for code repair tasks. The process involves generating a large set of candidate repair pairs (buggy code → fixed code) and then using a greedy algorithm to select a small, highly effective set of "golden pairs." This technique proves more compute-efficient and accurate than random sampling, and the selected examples generalize well to new problems within the same domain.