A3Bench - Overall: leaderboard

Zero-shot multiple-choice video QA benchmark (arXiv:2611.24762) measuring question-count-weighted accuracy (%) on aggressive action recognition and aggressor/victim/bystander role-action binding.

Metric: Overall Accuracy (%). Source: arxiv.org. Saturation forecast: Around June 2028. 14 models tracked. Updated 20 Nov 2026.

Top models

#ModelScoreOverall rank
1Qwen 3 VL 8B Instruct46.96#1035
2InternVL2.5-8B45.77#2370
3Ovis2.5-9B43.16#1189
4Gemma 4 26B A4B (IT)42.72#561
5InternVL3-9B42.55#1731
6GPT-5.141.3#242
7InternVL3.5-8B41.06#1599
8InternVideo2.5-8B40.7
9Qwen 3 VL 8B (Thinking)39.37#920
10VideoLLaMA3-7B38.22#2000
11Qwen 2.5 VL 7B Instruct37.3#1904
12ViCLIP23.82
13X-CLIP20.12
14LLaVA-Video-7B-Qwen218.5#1977

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=a3bench-overall · How It Works · Data refreshed daily, snapshot 2026-11-20.