A3Bench - Compound Questions: leaderboard

Zero-shot multiple-choice video QA accuracy (%) on multi-attribute compound role-action questions in A3Bench (arXiv:2611.24762).

Metric: Compound Questions Accuracy (%). Source: arxiv.org. Saturation forecast: Around 2029. 14 models tracked. Updated 20 Nov 2026.

Top models

#ModelScoreOverall rank
1Qwen 3 VL 8B Instruct46.65#1035
2InternVL2.5-8B45.01#2370
3Gemma 4 26B A4B (IT)39.65#561
4InternVideo2.5-8B39.34
5GPT-5.139.17#242
6InternVL3-9B39.06#1731
7Ovis2.5-9B37.99#1189
8Qwen 2.5 VL 7B Instruct37.15#1904
9InternVL3.5-8B37.03#1599
10Qwen 3 VL 8B (Thinking)36.41#920
11VideoLLaMA3-7B30.86#2000
12ViCLIP21.61
13LLaVA-Video-7B-Qwen219.34#1977
14X-CLIP15.75

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

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