EDIT preprint reports more efficient selection of AI responses for review

Researchers report in a preprint that EDIT, an algorithm for selecting AI responses for review, achieves a better trade-off between review cost and errors than direct language-model batch selection across four reasoning and factual question-answering benchmarks. The algorithm learns which interactions are likely to need corrections using feedback available only for interactions selected for review. It combines that estimate with a threshold controlling unnecessary reviews and a selection rule that accounts for cost. The researchers show that it respects the editor budget at every round. An editor can be a human expert or a more capable AI evaluation process; the reported comparison is limited to the four benchmarks.