StructVRM: Aligning Multimodal Reasoning with Structured and Verifiable Reward Models
Abstract
Existing Vision-Language Models often struggle with complex, multi-question reasoning tasks where partial correctness is crucial for effective learning. Traditional reward mechanisms, which provide a single binary score for an entire response, are too coarse to guide models through intricate problems with multiple sub-parts. To address this, we introduce StructVRM, a method that aligns multimodal reasoning with Structured and Verifiable Reward Models. At its core is a model-based verifier trained to provide fine-grained, sub-question-level feedback, assessing semantic and mathematical equivalence rather than relying on rigid string matching. This allows for nuanced, partial credit scoring in previously intractable problem formats. Extensive experiments demonstrate the effectiveness of StructVRM. Our trained model, Seed-StructVRM, achieves state-of-the-art performance on six out of twelve public multimodal benchmarks and our newly curated, high-difficulty STEM-Bench. The success of StructVRM validates that training with structured, verifiable rewards is a highly effective approach for advancing the capabilities of multimodal models in complex, real-world reasoning domains.
Cite
@article{arxiv.2508.05383,
title = {StructVRM: Aligning Multimodal Reasoning with Structured and Verifiable Reward Models},
author = {Xiangxiang Zhang and Jingxuan Wei and Donghong Zhong and Qi Chen and Caijun Jia and Cheng Tan and Jinming Gu and Xiaobo Qin and Zhiping Liu and Liang Hu and Tong Sun and Yuchen Wu and Zewei Sun and Chenwei Lou and Hua Zheng and Tianyang Zhan and Changbao Wang and Shuangzhi Wu and Zefa Lin and Chang Guo and Sihang Yuan and Riwei Chen and Shixiong Zhao and Yingping Zhang and Gaowei Wu and Bihui Yu and Jiahui Wu and Zhehui Zhao and Qianqian Liu and Ruofeng Tang and Xingyue Huang and Bing Zhao and Mengyang Zhang and Youqiang Zhou},
journal= {arXiv preprint arXiv:2508.05383},
year = {2025}
}