A Very Big Video Reasoning Suite
Maijunxian Wang, Ruisi Wang, Juyi Lin, Ran Ji, Thaddäus Wiedemer, Qingying Gao, Dezhi Luo, Yaoyao Qian, Lianyu Huang, Zelong Hong, Jiahui Ge, Qianli Ma, Hang He, Yifan Zhou, Lingzi Guo, Lantao Mei, Jiachen Li, Hanwen Xing, Tianqi Zhao, Fengyuan Yu, Weihang Xiao, Yizheng Jiao
2026-02-24
Summary
This paper introduces a new, very large dataset and evaluation tools designed to help computers better understand and reason about videos, going beyond just recognizing what's *in* a video to understanding *why* things happen in a video.
What's the problem?
Current video models are really good at making videos look realistic, but they aren't very good at actually understanding what's going on in the video and making logical connections. A big reason for this is that there just isn't enough data available to train these models to perform complex reasoning tasks about video content. It's hard to teach a computer to understand cause and effect, or how objects interact over time, without a lot of examples.
What's the solution?
The researchers created a massive dataset called VBVR, which includes over a million video clips covering 200 different reasoning tasks. They also built a way to test these models, called VBVR-Bench, that doesn't just rely on other computer programs to judge the answers, but uses rules and human-like scoring to make sure the evaluations are fair and understandable. They then used this dataset to see how well models improve as they are trained with more and more video data.
Why it matters?
This work is important because it provides the resources needed to push the field of video understanding forward. By having a large, well-organized dataset and a reliable way to test models, researchers can now focus on building AI that can truly 'think' about videos, not just 'see' them, which is a crucial step towards more intelligent and helpful AI systems.
Abstract
Rapid progress in video models has largely focused on visual quality, leaving their reasoning capabilities underexplored. Video reasoning grounds intelligence in spatiotemporally consistent visual environments that go beyond what text can naturally capture, enabling intuitive reasoning over spatiotemporal structure such as continuity, interaction, and causality. However, systematically studying video reasoning and its scaling behavior is hindered by the lack of large-scale training data. To address this gap, we introduce the Very Big Video Reasoning (VBVR) Dataset, an unprecedentedly large-scale resource spanning 200 curated reasoning tasks following a principled taxonomy and over one million video clips, approximately three orders of magnitude larger than existing datasets. We further present VBVR-Bench, a verifiable evaluation framework that moves beyond model-based judging by incorporating rule-based, human-aligned scorers, enabling reproducible and interpretable diagnosis of video reasoning capabilities. Leveraging the VBVR suite, we conduct one of the first large-scale scaling studies of video reasoning and observe early signs of emergent generalization to unseen reasoning tasks. Together, VBVR lays a foundation for the next stage of research in generalizable video reasoning. The data, benchmark toolkit, and models are publicly available at https://video-reason.com/ .