Aligning Language Model Benchmarks with Pairwise Preferences

Reinforcement Learning from Human Feedback has become the standard paradigm for language model alignment, where reward models directly determine alignment effectiveness. In this work, we focus on how to evaluate the generalizability of reward models. By "generalizability", we mean the ability of RMs to correctly rank responses to align with diverse user preferences. However, existing reward model benchmarks are typically designed around a universal preference, failing to assess this generalization. To address this critical gap, we introduce RMGAP, a benchmark comprising 1,097 instances across Chat, Writing, Reasoning, and Safety domains. Since different users exhibit diverse preferences for the same task, we first generate four distinct responses with different linguistic profiles for each collected prompt. However, the original prompt set lacks the specificity to convey different preferences. We therefore construct tailored prompts by contrasting these candidates and designing scenarios in which one response becomes the uniquely appropriate choice. Moreover, we observe that users often express the same preference using different phrasings, and thus extend each prompt with two paraphrased variants. Our evaluation of 24 state-of-the-art RMs reveals their substantial limitations: even the best RM achieves only 49.27% Best-of-N accuracy, highlighting considerable room for improvement in reward model generalization. Related data and code are available at https://github.com/nanzhi84/RMGAP.

With the rise in capabilities of large language models (LLMs) and their deployment in real-world tasks, evaluating LLM alignment with human preferences has become an important challenge. Current benchmarks average preferences across all users to compute aggregate ratings, overlooking individual user preferences when establishing model rankings. Since users have varying preferences in different contexts, we call for personalized LLM benchmarks that rank models according to individual needs. We compute personalized model rankings using ELO ratings and Bradley-Terry coefficients for 115 active Chatbot Arena users and analyze how user query characteristics (topics and writing style) relate to LLM ranking variations. We demonstrate that individual rankings of LLM models diverge dramatically from aggregate LLM rankings, with Bradley-Terry correlations averaging only $ρ= 0.04$ (57% of users show near-zero or negative correlation) and ELO ratings showing moderate correlation ($ρ= 0.43$). Through topic modeling and style analysis, we find users exhibit substantial heterogeneity in topical interests and communication styles, influencing their model preferences. We further show that a compact combination of topic and style features provides a useful feature space for predicting user-specific model rankings. Our results provide strong quantitative evidence that aggregate benchmarks fail to capture individual preferences for most users, and highlight the importance of developing personalized benchmarks that rank LLM models according to individual user preferences.

Selecting a pretrained language model, or evaluating a fine-tuned one, for a specific application is a high-value decision, yet the public benchmarks used to make it are poorly suited: a generic benchmark need not reflect a particular sub-domain or sub-task, and its scores are suspect when its items have leaked into pretraining and are recalled rather than solved. We present CoEval, an open framework that supplies a trustworthy, task-specific signal through ensemble self-evaluation: from a task or domain description, a pool of models rotates through all three roles, teacher, student, and judge, to generate a fresh, contamination-free benchmark, answer it, and score one another, with no human labels or raters. Because every model also answers as a student, the responses are the data that weight each question by its discriminative power and each judge by its consensus with the panel. Where ground truth exists, CoEval recovers the true ranking and tracks objective correctness at \r{ho}=0.86, and the weighting recovers the gold ranking of thirteen models at Spearman 0.95. Reliability comes from panel composition, not size: this label-free weighting zeroes out broken judges and down-weights saturated questions, so neither distorts the ranking. Generated items show zero verbatim overlap with five public benchmarks, the panel cancels verbosity bias and precludes same-family self-preference, and rankings are domain-specific: three different models top four de-novo domains, so a generic leaderboard misdirects most practitioners. The same pipeline reruns on each model release, giving any team a contamination-free leaderboard for its application.