Vision-language reward modeling faces a dilemma: generative approaches are interpretable but slow, while discriminative ones are efficient but act as opaque "black boxes." To bridge this gap, we propose VL-MDR (Vision-Language Multi-Dimensional Reward), a framework that dynamically decomposes evaluation into granular, interpretable dimensions. Instead of outputting a monolithic scalar, VL-MDR employs a visual-aware gating mechanism to identify relevant dimensions and adaptively weight them (e.g., Hallucination, Reasoning) for each specific input. To support this, we curate a dataset of 321k vision-language preference pairs annotated across 21 fine-grained dimensions. Extensive experiments show that VL-MDR consistently outperforms existing open-source reward models on benchmarks like VL-RewardBench. Furthermore, we show that VL-MDR-constructed preference pairs effectively enable DPO alignment to mitigate visual hallucinations and improve reliability, providing a scalable solution for VLM alignment.
@article{arxiv.2604.05445,
title = {Learning What Matters: Dynamic Dimension Selection and Aggregation for Interpretable Vision-Language Reward Modeling},
author = {Qiyuan Chen and Hongsen Huang and Jiahe Chen and Qian Shao and Jintai Chen and Hongxia Xu and Renjie Hua and Chuan Ren and Jian Wu},
journal= {arXiv preprint arXiv:2604.05445},
year = {2026}
}