MA-Bench (Closed-Ended): leaderboard
Metric: Accuracy (%) averaged over the six closed-ended tasks (coarse and fine micro-action recognition, single-action detail, multiple micro-action detail, micro-action sequence, part-proximity relation; multiple-choice and binary questions), MA-Bench's 1,000 micro-action videos (about 2 seconds each, 52 micro-action categories), 8 sampled frames, zero-shot, final answer without intermediate reasoning; higher is better. Source: arxiv.org. Saturation forecast: Around 2029. 23 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Flash | 50.7 | #237 |
| 2 | Qwen 2 VL 7B | 47.29 | #741 |
| 3 | GPT-4o | 44.87 | #333 |
| 4 | InternVL3-8B | 43.74 | #606 |
| 5 | Qwen 2.5 VL 7B Instruct | 43.71 | #643 |
| 6 | InternVL2-8B | 41.12 | #826 |
| 7 | Phi-4 Multimodal Instruct | 40.64 | #896 |
| 8 | Pixtral-12B | 40.08 | #795 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=ma-bench-closed-ended · How It Works · Data refreshed daily, snapshot 2026-10-11.