[diptanu]

This approach was studied in a paper called Evaporate+ - https://www.vldb.org/pvldb/vol17/p92-arora.pdf They used active learning to pick the best function among candidate functions generated by the LLM on a sampled set of data.