VGenST-Bench - Macro Average: leaderboard
Metric: Macro-averaged circular accuracy (%) on VGenST-Bench base multiple-choice questions; unweighted mean across the twelve task scores, each task score macro-averaged over applicable QA types; each model receives 8 uniformly sampled video frames; higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 15 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 3 Flash | 85.6 |
| 2 | GPT-5.4 | 82.7 |
| 3 | Gemini 3.1 Flash Lite | 79.6 |
| 4 | Gemma 4 31B (IT) | 72.5 |
| 5 | Kimi K2.6 | 71.7 |
| 6 | Qwen 3.5 27B | 68.7 |
| 7 | Gemma 4 26B A4B (IT) | 67.5 |
| 8 | GPT-5.4 Mini | 67 |
| 9 | Qwen 3.5 9B | 66.8 |
| 10 | Qwen 3.5 4B | 64 |
| 11 | GPT-5.4 Nano | 56.8 |
Interactive version: theaggregate.ai/benchmark?slug=vgenst-bench-macro-average · How It Works · Data refreshed daily, snapshot 2026-10-09.