TriViewBench - Tower Height Recovery: leaderboard
TriViewBench (2026) tests multi-view structural reasoning in multimodal models on synthetic 3D tower scenes shown as front, side and top views, with controlled object count and occlusion; models answer directly within 128 tokens. This board is accuracy in percent on the 1,923 Global Recovery questions that ask for the full sorted list of tower heights, scored by exact match of the list, across all complexity levels; higher is better.
Metric: Accuracy (%) over the 1,923 Global Recovery tower-height-list questions (recover the sorted height of every tower from the three views, exact match; the appendix table leaves out the hidden-count and occluded-reference questions that only Levels 2 and 4 carry), all four complexity levels; three orthographic views (front, side, top) of synthetic 3D tower scenes, direct prompting with a concise answer within 128 tokens, greedy decoding; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 18 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Flash | 75.51 | #453 |
| 2 | Claude 3.7 Sonnet | 56.42 | #458 |
| 3 | GPT-4o | 47.48 | #741 |
| 4 | Qwen 3 VL 32B Instruct | 40.77 | #585 |
| 5 | Qwen 3 VL 8B Instruct | 39.42 | #981 |
| 6 | Qwen 3 VL 4B Instruct | 38.85 | #1279 |
| 7 | Qwen 2.5 VL 32B Instruct | 19.19 | #1053 |
| 8 | Qwen 2.5 VL 7B Instruct | 11.44 | #1833 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=triviewbench-tower-height-recovery · How It Works · Data refreshed daily, snapshot 2026-11-13.