VSTAT - Atomic: leaderboard

VSTAT (Visual State Tracking benchmark, 2026) asks multimodal models 1,500 questions about 834 synthetic and real-world video clips that can only be answered by tracking a visual state through the whole video. This board is the score (%) on questions about atomic state structures; higher is better.

Metric: Score (%) on questions about atomic 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 2032. 23 models tracked.

Top models

#ModelScoreOverall rank
1Gemini 3.1 Pro (Preview) (High)40.1#94 (Gemini 3.1 Pro (Preview))
2Gemini 3.1 Pro (Preview) (Low)39.5#94 (Gemini 3.1 Pro (Preview))
3Molmo2-8B34.7#1483
4Gemini 3 Flash (Low)32.5#164 (Gemini 3 Flash)
5Qwen 3 VL 8B Instruct32.4#981
6Gemini 3 Flash (High)31.4#164 (Gemini 3 Flash)
7Qwen 3 VL 4B Instruct30.4#1279
8GLM-4.1V-9B (Thinking)26.9#1060 (GLM-4.1V-9B)
9Qwen 3 VL 8B (Thinking)26.8#887
10Qwen 3 VL 4B (Thinking)25.5#1296

No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.

Interactive version: theaggregate.ai/benchmark?slug=vstat-atomic · How It Works · Data refreshed daily, snapshot 2026-11-13.