Reproduce “VideoNexus: Video Perception Beyond the Reasoning Context” 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

Yancheng Long, Xuanyu Zheng, Wei Chen, Luoyuan Zhang, Haoyang Xu, Bin Wen, Tingting Gao

VideoNexus answers long-video questions by separating inspection from planner reasoning. Its perception calls locate and inspect intervals, then return compact, timestamped findings and selected frames rather than all inspected video. On LVBench, with 1,549 questions about 103 hour-scale videos, it reached 88.6% accuracy versus 85.0% for static ingestion; its lead over Gemini Agentic was not statistically significant, and the cost comparison is API-specific.

Summary by alphaXiv.

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