Abstract:Ranking LLMs via pairwise human feedback underpins current leaderboards for open-ended tasks, such as creative writing and problem-solving. We analyze ~89K comparisons in 116 languages from 52 LLMs from Arena, and show that the best-fit global Bradley-Terry (BT) ranking is misleading. Nearly 2/3 of the decisive votes cancel out, and even the top 50 models according to the global BT ranking are statistically indistinguishable (pairwise win probabilities are at most 0.53 within the top 50 models). We trace this failure to strong, structured heterogeneity of opinions across language, task, and time. Moreover, we find an important characteristic - *language* plays a key role. Grouping by language (and families) increases the agreement of votes massively, resulting in two orders of magnitude higher spread in the ELO scores (i.e., very consistent rankings). What appears as global noise is in fact a mixture of coherent but conflicting subpopulations.
To address such heterogeneity in supervised machine learning, we introduce the framework of $(\lambda, \nu)$-portfolios, which are small sets of models that achieve a prediction error at most $\lambda$, "covering" at least a $\nu$ fraction of users. We formulate this as a variant of the set cover problem and provide guarantees using the VC dimension of the underlying set system. On the Arena data, our algorithms recover just 5 distinct BT rankings that cover over 96% of votes at a modest $\lambda$, compared to the 21% coverage by the global ranking. We also provide a portfolio of 6 LLMs that cover twice as many votes as the top-6 LLMs from a global ranking. We further construct portfolios for a classification problem on the COMPAS dataset using an ensemble of fairness-regularized classification models and show that these portfolios can be used to detect blind spots in the data, which might be of independent interest to policymakers.
| Subjects: | Machine Learning (cs.LG); Discrete Mathematics (cs.DM); Emerging Technologies (cs.ET); Optimization and Control (math.OC) |
| ACM classes: | I.2.6; I.2.7; H.3.3; K.4 |
| Cite as: | arXiv:2605.06656 [cs.LG] |
| (or arXiv:2605.06656v1 [cs.LG] for this version) | |
| https://doi.org/10.48550/arXiv.2605.06656 arXiv-issued DOI via DataCite (pending registration) |
From: Swati Gupta [view email]
[v1]
Thu, 7 May 2026 17:57:58 UTC (1,046 KB)
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。