HAVEN - Grounding: leaderboard
Multimodal grounding on HAVEN (2026), a video understanding benchmark with aligned frames, shots, descriptions and full video documents: the model links each summary sentence to the video shots it describes. The score is F1 between predicted and annotated shots in percent; higher is better.
Metric: Multimodal grounding F1 (%) between predicted and ground-truth shots for each summary sentence, zero-shot on a random sample of 100 videos per source dataset; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 6 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.2 | 51 | #185 |
| 2 | GLM-4.6V-Flash | 45.7 | #1040 |
| 3 | Qwen 3 VL 8B Instruct | 44.3 | #1029 |
| 4 | InternVideo2.5-Chat-8B | 39.4 | #2036 |
| 5 | Qwen 2.5 VL 7B Instruct | 29.9 | #1902 |
| 6 | VideoLLaMA3-7B | 25.3 | #2050 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=haven-grounding · How It Works · Data refreshed daily, snapshot 2026-11-21.