MOV-Bench (AOP-Agent): leaderboard
MOV-Bench (Xu et al., 2026) has 519 multiple-choice questions on FineVideo clips that need two to four hops of reasoning over evidence spread across several segments and across both the audio and the visual stream; questions answerable from language alone were filtered out and the rest human-verified. This board runs each omni-modal model inside the authors' AOP-Agent framework (a hierarchical audio-visual memory of the video and an observe-reflect-replan loop of agents sharing that backbone); accuracy in percent, higher is better.
Metric: Accuracy (%; 519 audio-visual multi-hop multiple-choice questions; the AOP-Agent framework, hierarchical omni-modal memory and an observe-reflect-replan loop with every agent on the listed backbone). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 3 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen3-Omni-30B-A3B-Instruct (AOP-Agent) | 62.62 | |
| 2 | Qwen3-Omni-30B-A3B-Thinking (AOP-Agent) | 62.24 | |
| 3 | Qwen2.5-Omni-7B (AOP-Agent) | 47.01 |
Interactive version: theaggregate.ai/benchmark?slug=mov-bench-aop-agent · How It Works · Data refreshed daily, snapshot 2026-11-12.