VideoZeroBench - Spatial Grounding: leaderboard
VideoZeroBench (2026) asks 500 manually annotated, deliberately hard open-ended questions with exact short answers about 138 long videos in 13 domains, each with annotated temporal intervals and bounding boxes as supporting evidence. Mean visual IoU of the boxes the model draws at given key-frame timestamps against the annotated boxes, in percent; higher is better.
Metric: Mean visual IoU (%) of the bounding boxes the model predicts at the given key-frame timestamps against the annotated boxes, averaged over frames (0 for a missing box or a question without spatial evidence), on VideoZeroBench's 500 manually annotated open-ended questions (exact short answers) over 138 long videos in 13 domains; Gemini models get the raw video through the API, the others 1 fps frames up to the printed cap, temperature 0; higher is better. Source: arxiv.org. Saturation forecast: Around August 2027. 17 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Seed 2.0 Pro | 21.8 | #222 |
| 2 | GPT-5.2 | 12.6 | #185 |
| 3 | Open-o3-Video-4B | 9.9 | |
| 4 | Gemini 3 Pro | 9.6 | #145 |
| 5 | Gemini 2.5 Pro | 8.6 | #276 |
| 6 | Qwen 3.5 397B A17B | 8 | #264 |
| 7 | Video-R1 | 5.5 | #1379 |
| 8 | Gemini 2.5 Flash | 4.9 | #463 |
| 9 | VideoRFT-7B | 4.2 | #1863 |
| 10 | Qwen 3 VL 235B A22B | 3.6 | #485 |
| 11 | VideoChat-R1.5-7B | 2.9 | #1254 |
| 12 | Qwen 3 VL 8B | 2.4 | #917 |
| 13 | Qwen 2.5 VL 7B | 1.4 | #1862 |
| 14 | InternVL3.5-4B | 1.2 | #1930 |
| 15 | Qwen 3 VL 4B | 1.2 | #1151 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=videozerobench-spatial-grounding · How It Works · Data refreshed daily, snapshot 2026-11-21.