LLM Judge Validation Under Sparse Overlap: From Inference to Design
It is proved that overlap sparsity is the first-order determinant of wrong deployment decisions: at 5% pairwise overlap, wrong-decision rates reach 25% and the probability of selecting the wrong best judge among ten candidates is 65%.
Jun-Xuan Li, Arko Provo Mukherjee, Soumyabrata Pal
· 0 citations