VSTAT - Set: leaderboard
Metric: Score (%) on questions about set state structures: 1,500 questions on 834 synthetic (Blender) and real-world video clips that require tracking a visual state through the whole video; accuracy on multiple-choice questions and mean relative accuracy on numerical questions (%), open models at their best of 16 to 128 uniformly sampled frames; higher is better. Source: arxiv.org. Saturation forecast: Around 2031. 23 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) (Low) | 51.9 |
| 2 | Gemini 3.1 Pro (Preview) (High) | 50 |
| 3 | Gemini 3 Flash (High) | 48.4 |
| 4 | InternVL3.5-8B | 41.8 |
| 5 | GLM-4.1V-9B (Thinking) | 40.8 |
| 6 | Qwen 3 VL 4B Instruct | 39.8 |
| 7 | Molmo2-8B | 39.1 |
| 8 | Qwen 3 VL 8B Instruct | 37.9 |
| 9 | Qwen 3 VL 4B (Thinking) | 29.5 |
| 10 | Qwen 3 VL 8B (Thinking) | 28.5 |
Interactive version: theaggregate.ai/benchmark?slug=vstat-set · How It Works · Data refreshed daily, snapshot 2026-09-29.