VidOmni-Bench: leaderboard
VidOmni-Bench (Kim et al., 2026) tests fine-grained video understanding by asking a Video-LLM to verify, sentence by sentence, whether each event in a dense caption of a video is supported by the video; human-labelled incorrect sentences from model-written captions serve as hard negatives. F1 in percent for detecting the incorrect sentences, averaged over video-caption pairs, with video-only input; higher is better.
Metric: F1 (%; sentence-wise event verification: for each of the dense captions written by five Video-LLMs for the 500 videos (4 s to 90 min, five complexity types), the model labels every sentence Correct or Incorrect and the human-labelled Incorrect sentences are the detection target; precision, recall and F1 are computed per video-caption pair and averaged over pairs; video-only input). Source: arxiv.org. Saturation forecast: Around April 2027. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash | 39.8 | #159 |
| 2 | GPT-5 Mini | 28.9 | #349 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=vidomni-bench · How It Works · Data refreshed daily, snapshot 2026-11-12.