Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Zhang, Xiangxiang, Wei, Jingxuan, Zhong, Donghong, Chen, Qi, Jia, Caijun, Tan, Cheng, Gu, Jinming, Qin, Xiaobo, Liu, Zhiping, Hu, Liang, Sun, Tong, Wu, Yuchen, Sun, Zewei, Lou, Chenwei, Zheng, Hua, Zhan, Tianyang, Wang, Changbao, Wu, Shuangzhi, Lin, Zefa, Guo, Chang, Yuan, Sihang, Chen, Riwei, Zhao, Shixiong, Zhang, Yingping, Wu, Gaowei, Yu, Bihui, Wu, Jiahui, Zhao, Zhehui, Liu, Qianqian, Tang, Ruofeng, Huang, Xingyue, Zhao, Bing, Zhang, Mengyang, Zhou, Youqiang
Format: Preprint
Veröffentlicht: 2025
Schlagworte:
Online-Zugang:https://arxiv.org/abs/2508.05383
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866915433771171840
author Zhang, Xiangxiang
Wei, Jingxuan
Zhong, Donghong
Chen, Qi
Jia, Caijun
Tan, Cheng
Gu, Jinming
Qin, Xiaobo
Liu, Zhiping
Hu, Liang
Sun, Tong
Wu, Yuchen
Sun, Zewei
Lou, Chenwei
Zheng, Hua
Zhan, Tianyang
Wang, Changbao
Wu, Shuangzhi
Lin, Zefa
Guo, Chang
Yuan, Sihang
Chen, Riwei
Zhao, Shixiong
Zhang, Yingping
Wu, Gaowei
Yu, Bihui
Wu, Jiahui
Zhao, Zhehui
Liu, Qianqian
Tang, Ruofeng
Huang, Xingyue
Zhao, Bing
Zhang, Mengyang
Zhou, Youqiang
author_facet Zhang, Xiangxiang
Wei, Jingxuan
Zhong, Donghong
Chen, Qi
Jia, Caijun
Tan, Cheng
Gu, Jinming
Qin, Xiaobo
Liu, Zhiping
Hu, Liang
Sun, Tong
Wu, Yuchen
Sun, Zewei
Lou, Chenwei
Zheng, Hua
Zhan, Tianyang
Wang, Changbao
Wu, Shuangzhi
Lin, Zefa
Guo, Chang
Yuan, Sihang
Chen, Riwei
Zhao, Shixiong
Zhang, Yingping
Wu, Gaowei
Yu, Bihui
Wu, Jiahui
Zhao, Zhehui
Liu, Qianqian
Tang, Ruofeng
Huang, Xingyue
Zhao, Bing
Zhang, Mengyang
Zhou, Youqiang
contents 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.
format Preprint
id arxiv_https___arxiv_org_abs_2508_05383
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle StructVRM: Aligning Multimodal Reasoning with Structured and Verifiable Reward Models
Zhang, Xiangxiang
Wei, Jingxuan
Zhong, Donghong
Chen, Qi
Jia, Caijun
Tan, Cheng
Gu, Jinming
Qin, Xiaobo
Liu, Zhiping
Hu, Liang
Sun, Tong
Wu, Yuchen
Sun, Zewei
Lou, Chenwei
Zheng, Hua
Zhan, Tianyang
Wang, Changbao
Wu, Shuangzhi
Lin, Zefa
Guo, Chang
Yuan, Sihang
Chen, Riwei
Zhao, Shixiong
Zhang, Yingping
Wu, Gaowei
Yu, Bihui
Wu, Jiahui
Zhao, Zhehui
Liu, Qianqian
Tang, Ruofeng
Huang, Xingyue
Zhao, Bing
Zhang, Mengyang
Zhou, Youqiang
Artificial Intelligence
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.
title StructVRM: Aligning Multimodal Reasoning with Structured and Verifiable Reward Models
topic Artificial Intelligence
url https://arxiv.org/abs/2508.05383