RMGAP: Benchmarking the Generalization of Reward Models across Diverse Preferences

2026-05-03

AI Summary

Existing reward model benchmarks predominantly rely on uniform human preferences, limiting their ability to evaluate generalization under diverse user preferences. This work proposes RMGAP—the first evaluation framework specifically designed to assess reward model generalization across heterogeneous preferences. RMGAP systematically measures ranking generalization by generating four linguistically distinct responses per task and incorporating preference-specific prompts, semantically equivalent rewrites, and multi-response comparison scenarios. The authors evaluate 24 state-of-the-art reward models on 1,097 instances spanning dialogue, writing, reasoning, and safety domains. Results reveal that even the best-performing model achieves only a 49.27% Best-of-N accuracy, highlighting a significant gap in current models’ capacity to generalize across varied preference distributions.

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.

Problem

• reward model
• generalization
• human preferences
• benchmark
• alignment

Innovation

• reward model generalization
• preference diversity
• RMGAP benchmark
• human feedback alignment
• Best-of-N accuracy