Pair Difficulty Matters: Rethinking Pairwise LLM-as-a-Judge Evaluation and Consistency

A new arXiv paper (2609.37577v1) argues that the three standard proxies for evaluating pairwise LLM-as-a-judge reliability — position bias, transitivity, and pairwise agreement — are misleading because they are dominated by close-rank-gap pairs where inconsistency is information-theoretically expected. The authors formalize this under Bradley-Terry geometry and validate it in a controlled simulation and on two human-rated corpora, finding the proxies correlate only weakly with ranking accuracy against gold and that their predictive component concentrates in the far-gap regime. They recommend assessing judges on rank-gap-conditional metrics, ideally against human rankings, with code released at github.com/brunobrocai/PairDifficulty.

arXiv:2609.37577v1 Announce Type: new Abstract: Large Language Model judges are widely used to rank texts and text-generating systems through pairwise comparison, and their reliability is typically assessed via three proxies: position bias, transitivity, and pairwise agreement self- or human-labeled . Because these proxies drive judge selection and benchmarking, a substantial literature reporting that judges perform poorly on them risks steering practitioners away from otherwise capable evaluators. We argue this assessment is misleading. Under the Bradley--Terry geometry underlying pairwise aggregation, each proxy is dominated by close-rank-gap pairs, where inconsistency is information-theoretically expected and individual verdicts contribute little to the aggregate ranking; far-gap pairs carry the ranking signal but barely move the proxies. We formalize this argument and validate it in a controlled simulation and on two human-rated corpora: the proxies correlate only weakly with ranking accuracy against gold, and their predictive component concentrates in the far-gap regime. Judges should therefore be assessed on rank-gap-conditional metrics, ideally against human rankings. Code at https://github.com/brunobrocai/PairDifficulty.