metadata
license: mit
task_categories:
- video-text-to-text
Video-Browsecomp
Project Page | Paper | Code
Video-Browsecomp is a benchmark designed for evaluating open-ended agentic browsing tasks that enforce a mandatory dependency on videos. It was introduced in the paper "Video-Browser: Towards Agentic Open-web Video Browsing".
The benchmark challenges autonomous agents to research, watch, and analyze video content to answer complex, open-ended user queries. Along with the benchmark, the authors propose Video-Browser, a novel agent leveraging Pyramidal Perception to efficiently filter information using metadata and perform fine-grained visual verification only when necessary.
Citation
If you use the Video-Browsecomp benchmark or the Video-Browser agent in your research, please cite:
@misc{liang2026videobrowseragenticopenwebvideo,
title={Video-Browser: Towards Agentic Open-web Video Browsing},
author={Zhengyang Liang and Yan Shu and Xiangrui Liu and Minghao Qin and Kaixin Liang and Paolo Rota and Nicu Sebe and Zheng Liu and Lizi Liao},
year={2026},
eprint={2512.23044},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2512.23044},
}