STRAND: leaderboard
STRAND (Nguyen et al., 2026) tests object-centric spatio-temporal monitoring in video with human-verified yes-or-no questions about object identities, states, actions and interactions over time, each broken into the prerequisite sub-questions it depends on. Faithful accuracy in percent credits a target question only when the model also answers all of its sub-questions correctly, over 977 targets on 88 videos; higher is better.
Metric: Faithful accuracy (%; share of the 977 target yes/no questions over 88 videos for which the model answers the target and every prerequisite sub-question correctly, computed over all targets; 2,516 sub-questions; end-to-end video input; majority-class predictor 19.1). Source: arxiv.org. Saturation forecast: Not forecast. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 VL 32B (Thinking) | 46.1 | #557 |
| 2 | Qwen 3.5 27B | 45.5 | #358 |
| 3 | Gemini 3.1 Pro (Preview) | 38.9 | #91 |
| 4 | Claude Sonnet 4.6 | 30.2 | #158 |
| 5 | InternVL3-78B | 29.7 | #733 |
| 6 | Gemini 3 Flash | 28.7 | #159 |
| 7 | GPT-5 | 20 | #163 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=strand · How It Works · Data refreshed daily, snapshot 2026-11-12.