Reproduce “VideoJev: System One Model for Video Understanding and Interaction” 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
Yuhan Zhu, Xiangyu Zeng, Qingyi Si, Chenxu Yang, Yifei Li, Luozheng Qin, Junhao Zhou, Dingyu Yao and 10 more
VideoJev scores candidate time intervals to locate events and candidate answers for video questions, reusing one video prefix across queries through cached or joint inference. It reports 65.3 mean temporal intersection-over-union, versus 58.6 for its 4B backbone, and a four-benchmark QA mean of 74.3, near 74.9 for Qwen3.5-4B Thinking; evaluations cover grounding and multiple-choice QA, not agent workflows.
Key insights
- Temporal grounding and multiple-choice QA can be framed as scoring candidates through one shared interface, rather than generating a text response for each task.
- Reusing video computation supports repeated decisions about the same visual input, and joint training across grounding and QA improves average performance on both tasks.
- Direct scores provide useful ranking signals but are not fully calibrated: both model sizes tend to overestimate QA correctness.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.