VideoJev scores temporal intervals and textual options directly from shared video representations, unifying grounding and multiple-choice question answering without word-by-word token generation.
Yuhan Zhu and 17 collaborators release checkpoints and code that score time intervals and answers against video frames, reusing stored video calculations across questions. Author-reported speedups measure H200 forward passes, excluding preprocessing and transfers, not complete app latency. Question-answering tests use no audio or subtitles, and training includes an internal question-answer corpus. Reported confidence calibration errors are 8.65 and 8.13 percentage points; confident answers can still be wrong.
How you can use it
Start from the released checkpoints and examples if you need to find a described event in a video or choose an answer about it. A developer can reuse the video’s stored calculations for later questions, or ask several together. The reported tests use frames without audio or subtitles, and confident answers can still be wrong.
