- The problem
- Video assistants may ask many questions about the same recording, such as when something happened and which answer best fits what was seen. Generating each answer word by word is slow, especially when the answer is only a time range or a choice from a list.
- The idea
- VideoJev turns these into direct comparisons: it scores possible time ranges or written answers against the video, then picks the best-supported choice or joins the strongest neighboring time ranges. It also keeps the video’s processed information available, so later questions can reuse it—or several questions can be answered together.
- What it got
- On video event timing, its two model sizes improve average overlap with the correct time ranges by 6.7 and 10.8 points over their starting models, while running about 2.8–3.1 times faster. For multiple-choice video questions, it matches reasoning-enabled models while running more than 100 times faster.