Augmenting Human Evaluation with LLM Judges: How Many Human Reviews Do You Need?

In three linesarXiv paper proposing formal framework for combining LLM and human evaluations. Uses doubly robust estimator (missing data literature) to determine optimal number of human reviews needed. Shifts LLM role from substitutive to auxiliary in two-stage sampling design.

Summary generated by Claude — human-verified