A Judge-Aware Ranking Framework for Evaluating Large Language Models without Ground Truth
This addresses the issue of misleading evaluations in open-ended LLM tasks for researchers and practitioners, though it is incremental as it builds on existing ranking models.
The paper tackles the problem of biased and unreliable leaderboards in evaluating large language models (LLMs) without ground truth, caused by treating all judge LLMs equally, and proposes a judge-aware ranking framework that improves agreement with human preferences and achieves higher data efficiency.
Evaluating large language models (LLMs) on open-ended tasks without ground-truth labels is increasingly done via the LLM-as-a-judge paradigm. A critical but under-modeled issue is that judge LLMs differ substantially in reliability; treating all judges equally can yield biased leaderboards and misleading uncertainty estimates. More data can make evaluation more confidently wrong under misspecified aggregation. We propose a judge-aware ranking framework that extends the Bradley-Terry-Luce model by introducing judge-specific discrimination parameters, jointly estimating latent model quality and judge reliability from pairwise comparisons without reference labels. We establish identifiability up to natural normalizations and prove consistency and asymptotic normality of the maximum likelihood estimator, enabling confidence intervals for score differences and rank comparisons. Across multiple public benchmarks and a newly collected dataset, our method improves agreement with human preferences, achieves higher data efficiency than unweighted baselines, and produces calibrated uncertainty quantification for LLM rankings.