Video-Browser: Towards Agentic Open-web Video Browsing

Zhengyang Liang1,♣, Yan Shu2,♣, Xiangrui Liu3, Minghao Qin3, Nicu Sebe2, Zheng Liu4, Lizi Liao1
1Singapore Management University, 2University of Trento,
3Beijing Academy of Artificial Intelligence, 4Hong Kong Polytechnic University
Paper Benchmark Code

Abstract

The evolution of autonomous agents is redefining information seeking, transitioning from passive retrieval to proactive, open-ended web research. However, while textual and static multimodal agents have seen rapid progress, a significant modality gap remains in the web's most dynamic modality: video. Existing video benchmarks predominantly focus on passive perception, feeding curated clips to models without requiring external retrieval. They fail to evaluate agentic video research, which necessitates actively interrogating video timelines, cross-referencing dispersed evidence, and verifying claims against the open web.

To bridge this gap, we present Video-BrowseComp, a challenging benchmark comprising 210 questions tailored for open-web agentic video reasoning. Unlike prior benchmarks, Video-BrowseComp enforces a mandatory dependency on temporal visual evidence, ensuring that answers cannot be derived solely through text search but require navigating video timelines to verify external claims. Our evaluation of state-of-the-art models reveals a critical bottleneck: even advanced search-augmented models like GPT-5.1 (w/ Search) achieve only 15.24% accuracy. Our analysis reveals that these models largely rely on textual proxies, excelling in metadata-rich domains (e.g., TV shows with plot summaries) but collapsing in metadata-sparse, dynamic environments (e.g., sports, gameplay) where visual grounding is essential. As the first open-web video research benchmark, Video-BrowseComp advances the field beyond passive perception toward proactive video reasoning.

Video-Browser Agent

To tackle the challenges of open-web agentic video research, we propose Video-Browser, an agentic framework composed of three specialized modules sharing a global memory:

  • Planner: Acts as the cognitive controller, decomposing user queries into sequential sub-tasks and generating search queries based on interaction history.
  • Watcher: Functions as a high-efficiency filter and localizer using a Pyramidal Perception mechanism. It progressively prunes the search space through Semantic Filtering (Stage I), Sparse Localization (Stage II), and Zoom-in (Stage III) to identify and decode relevant visual evidence without overwhelming computational costs.
  • Analyst: Synthesizes the final answer by reasoning over the accumulated multimodal evidence and the query context.

This architecture enables Video-Browser to actively interrogate video timelines and cross-reference dispersed evidence, achieving a 71.8% relative improvement over state-of-the-art search-augmented models (26.19% vs 15.24% accuracy).

Examples

Video-BrowseComp Examples

Figure 1: Examples of questions and tasks in Video-BrowseComp.

Dataset Statistics

Video-BrowseComp Statistics

Figure 2: Distribution of video categories and difficulty levels.

Leaderboard

We evaluate state-of-the-art models on Video-BrowseComp. The accuracy (%) is reported for Overall (OA) and three difficulty levels (Level 1, Level 2, Level 3).

Filter:
Model Overall Acc (%) Level 1 (%) Level 2 (%) Level 3 (%) Calibration Error (%)

Submission Guide

We welcome submissions of new models to the Video-BrowseComp leaderboard! To submit your results, please follow these steps:

  1. Run your model on the Video-BrowseComp benchmark available on Hugging Face.
  2. Prepare a JSON file containing your model's predictions and the calculated accuracy for each level.
  3. Email your results, including model details (name, size, type) and a link to your paper or project page, to our team.

Contact us for submission:

Citation

@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 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}, 
}