VSTAT - Sequence: leaderboard

Metric: Score (%) on questions about sequence 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 April 2028. 23 models tracked.

Top models

#ModelScore
1Gemini 3 Flash (High)62.4
2Gemini 3.1 Pro (Preview) (Low)60.8
3Gemini 3.1 Pro (Preview) (High)56.8
4Molmo2-8B36.3
5InternVL3.5-8B33.8
6Qwen 3 VL 8B Instruct33.3
7GLM-4.1V-9B (Thinking)33.2
8Qwen 3 VL 4B Instruct32.8
9Qwen 3 VL 8B (Thinking)32.7
10Qwen 3 VL 4B (Thinking)30.8

Interactive version: theaggregate.ai/benchmark?slug=vstat-sequence · How It Works · Data refreshed daily, snapshot 2026-09-29.