Pairwise Reference Alignment as a Model-Level Ordinal Observable

In three linesTheoretical paper defining pairwise reference alignment as an ordinal observable for language model evaluation. Formulates statistical framework to measure whether a model ranks preferred responses above rejected responses, with finite-sample estimators and concentration bounds. Empirical validation on Qwen2.5 and RewardBench.

Summary generated by Claude — human-verified