LLM Evaluation as Tensor Completion: Low Rank Structure and Semiparametric Efficiency
A score-whitening method that equalizes local Fisher information and restores stable inference at the optimal sample-complexity scale is introduced that provides a principled framework for uncertainty quantification in LLM evaluation and more broadly for inference on low-rank structures from pairwise data.