Reproduce “VideoSeek: Long-Horizon Video Agent with Tool-Guided Seeking” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Jingyang Lin, Jialian Wu, Jiang Liu, Ximeng Sun, Ze Wang, Xiaodong Yu, Jiebo Luo, Zicheng Liu and 1 more

VideoSeek answers long-video questions by iteratively using the question and accumulated observations to select overview, skim, or focus tools, rather than densely describing the entire video. On LVBench with subtitles, it reaches 76.7% accuracy while processing 27.2 frames on average, versus GPT-5’s 66.5% and 384 frames. Without subtitles, it is not the top-scoring system.

Key insights

  • Active, tool-guided evidence seeking, inspired by human video consumption patterns, enables more efficient utilization of visual information for long-horizon tasks compared to dense preprocessing.
  • Leveraging the video's underlying logic flow (e.g., explicit textual storylines from subtitles) dramatically reduces the necessary frame budget and enhances performance by localizing critical segments.
  • A tiered toolkit providing multi-granular interaction with video content is crucial for balancing global context identification with precise, fine-grained information extraction.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.