VideoZeroBench - Level-4 (Answer with Temporal 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. Level-4 accuracy: a question counts only when the model answers it correctly and its predicted evidence intervals reach a temporal IoU above 0.3; in percent of all questions, higher is better.

Metric: Level-4 accuracy (%), the share of all 500 questions answered correctly at Level-3 whose predicted evidence intervals reach temporal IoU above 0.3, 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 2029. 17 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Gemini 3 Pro8#146
2Gemini 2.5 Pro7.2#280
3Gemini 2.5 Flash5.4#466
4Seed 2.0 Pro3.8#226
5Qwen 3 VL 235B A22B3.4#486
6GPT-5.21.8#190
7Qwen 3 VL 4B1.4#1163
8Qwen 3 VL 8B0.6#927
9Open-o3-Video-4B0.4
10InternVL3.5-8B0.2#1618
11Qwen 3.5 397B A17B0#270
12Qwen 2.5 VL 7B0#1878
13InternVL3.5-4B0#1946
14Video-R10#1384
15VideoChat-R1.5-7B0#1260

No result here: #2 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=videozerobench-level-4-answer-with-temporal-grounding · How It Works · Data refreshed daily, snapshot 2026-11-22.