Preference Leakage

A bias in LLM-as-a-Judge evaluation where the judge model favors outputs from models it is related to, for example ones trained on similar data or by the same lab, skewing benchmark comparisons in ways that are hard to detect without checking for it directly.