Ranking Large Language Models with Human Preferences: A Game-Theoretic and Bayesian Comparative Study

Abstract

The rapid growth of large language models (LLMs) has created a demand for reliable, interpretable, and scalable ranking systems capable of comparing models based on human preferences. Traditional benchmarks often fail to capture qualitative nuances in open-domain dialog, where subjective judgments play a central role. In this work, we conduct a comparative study of six prominent ranking algorithms Elo, Glicko, TrueSkill, Bradley-Terry, Markov Chain-based ranking, and a novel HawasRank algorithm applied to the Chatbot Arena dataset containing 244,978 pairwise human preference comparisons. HawasRank is a divergence-based, game-theoretic ranking method inspired by Bregman optimization frameworks, designed to improve convergence speed, transitivity preservation, and computational efficiency in human-in-the-loop LLM evaluations. In this study, we carefully compare these algorithms based on several important criteria, such as predictive accuracy, the occurrence of transitivity violations, sensitivity to hyperparameters, convergence behavior, and CPU resource consumption. The findings of our analysis reveal the trade-offs that exist between different ranking approaches and offer practical guidance for choosing appropriate algorithms, especially in large-scale and evolving LLM evaluation environments.