Tag
This paper proposes a two-stage sampling design where LLM evaluations are used to augment, rather than replace, human ratings, and provides guidance on determining sample sizes for human and LLM reviews using a doubly robust estimator from missing data literature.