RMGAP: Benchmarking the Generalization of Reward Models across Diverse Preferences

TL;DR

This paper introduces RMGAP, a benchmark for evaluating how well reward models generalize across diverse user preferences in language tasks, revealing significant limitations of current models.

Contribution

The paper presents RMGAP, a novel benchmark with 1,097 instances across multiple domains, designed to assess reward model generalization to diverse preferences, and provides an evaluation of 24 state-of-the-art RMs.

Findings

Best reward model achieves only 49.27% accuracy in ranking responses.

Current reward models show substantial limitations in generalizing to diverse preferences.

The benchmark reveals significant room for improvement in reward model robustness.

Abstract

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.