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

In three linesarXiv paper proposing a formal framework for combining LLM and human evaluations. Uses a doubly robust estimator (missing data approach) to determine optimal sample sizes of human ratings needed for benchmark validation, shifting LLMs from substitutive to auxiliary role.

Summary generated by Claude — human-verified