Wednesday, 14 October 2026 · 7 min read
VideoJev unifies video event localization and QA with efficient direct scoring
VideoJev adapts pretrained multimodal LLMs to directly score temporal intervals and answer options for video understanding, improving accuracy by up to 10.8 mIoU points and speeding inference up to 100× on multiple benchmarks.
Based on the paper VideoJev: System One Model for Video Understanding and Interaction by Yuhan Zhu, Xiangyu Zeng, Qingyi Si, Chenxu Yang and 14 others
Challenges in timely video understanding and decision-making
Video understanding for visual agents requires not only recognizing content but making timely, structured decisions to track progress and guide next actions. For example, an assistant following a cooking video must verify if an ingredient was added before proceeding. This demands efficient reuse of visual context across multiple queries about the same video.
Existing models either make efficient but task-specific predictions or use multimodal large language models (MLLMs) that generate answers autoregressively, which can be slow and less direct for structured decisions like temporal grounding or multiple-choice QA.
The System One paradigm proposes making direct, typed probabilistic judgments without token-by-token generation, but extending this to video requires handling temporal structure and sharing expensive visual computation across queries efficiently.
Sign in to read the whole paper
You have read the opening. The rest is free with a Bezen account, and no card is needed.
In the rest of this paper
- How the method works, step by step, with a flow chart
- Three key formulas from the paper, explained symbol by symbol
- Three figures from the paper, and how to read them
- How it was tested, what was found, and where it falls short