SSMNBench - Multi-View Necessity: leaderboard
SSMNBench (2026) tests multimodal models on crowded, occlusion-heavy multi-view scenes of people, separating questions that one view can answer from questions that need evidence fused across views. This board is mean accuracy in percent over the six multi-view-necessity tasks, given the annotated necessary views and four answer options; higher is better.
Metric: Mean accuracy (%) over the six multi-view-necessity tasks (distinct human counting, counting with attribute, relative cross-person pose, relative distance, global orientation, global depth ordering), where no single view suffices; four-option multiple choice on dense, occlusion-heavy multi-view scenes of people, given exactly the annotated necessary views (the +0 setting); 300 questions per task, images resized to 1920x1080, single-letter answers. Source: arxiv.org. Saturation forecast: Around September 2028. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro | 44 | #266 |
| 2 | GPT-5.2 | 42.4 | #176 |
| 3 | Gemini 2.5 Flash | 39.4 | #453 |
| 4 | InternVL3-78B | 39.3 | #738 |
| 5 | Qwen 2.5 VL 32B Instruct | 38.9 | #1053 |
| 6 | Qwen 2.5 VL 72B Instruct | 38.3 | #846 |
| 7 | Qwen 2.5 VL 7B Instruct | 35.3 | #1833 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=ssmnbench-multi-view-necessity · How It Works · Data refreshed daily, snapshot 2026-11-13.