Statistically Reliable LLM-Based Ranking Evaluation via Prediction-Powered Inference

Researchers have developed PRECISE, a method that combines small human-labeled datasets with large LLM-generated judgments to produce bias-corrected estimates of ranking evaluation metrics. The approach reduces standard error by 21% on the ESCI benchmark and correctly identified the best system variant in a production environment, confirmed by A/B testing showing a +407 basis point increase in daily sales.

arXiv:2606.05308v1 Announce Type: new Abstract: With PRECISE, we extended Prediction-Powered Inference to produce bias-corrected estimates of ranking evaluation metrics by combining a small human-labeled set with a large LLM-judged set. PPI is provably unbiased regardless of the LLM judge's error profile. We make it applicable to hierarchical metrics like Precision@K, where annotations are per-document but the metric is per-query, by reducing the output-space computation from O 2^|C| to O 2^K . On the ESCI benchmark, augmenting 30 human annotations with Claude 3 Sonnet judgments reduces the standard error of Precision@4 estimates from 4.45 to 3.50 a 21% relative reduction . In a production system, our framework correctly identified the best of three system variants from 100 human labels and 2 hours of domain-expert annotation; A/B testing confirmed this ranking with +407 bps in daily sales.