MARS: Margin and Semantic-Aware Data Augmentation for Reward Modeling

Abstract

Reward modeling is central to RLHF, RLAIF, and PPO-based alignment, but its reliability is often limited by scarce and heterogeneous human preference data. In this paper, we introduce MARS (Margin and Semantic-Aware Data Augmentation for Reward Modeling), an adaptive augmentation framework for controlled low-resource reward modeling. MARS allocates more augmentation to low-margin preference pairs and uses semantic-distance-based refinement to improve chosen-rejected contrast before generating synthetic preference samples. Across three preference datasets, two reward-model backbones, and downstream alignment evaluations, MARS improves average RewardBench performance and alignment win rates over uniform augmentation, WoN, and AdaBoost-style baselines. Ablations and independent-judge evaluations suggest that the gains are not solely explained by semantic refinement alone or GPT-4.1 judge coupling.