3D Scene Structure - Multi-Object Removal: leaderboard
Spatial counterfactual task from the Unnamed 3D-scene benchmark from 'Do Vision-Language Models Understand 3D Scenes or Just Catalogue Objects?' (Maheshwari, Sahu and Verma, 2026): the model names the smallest set of objects, in removal order, whose removal would reveal a given hidden object in a synthetic indoor scene. Trained annotators grade every response with no partial credit; models run with thinking enabled at temperature 0. Correctness in percent over 330 scenes, credited only for the exact set; higher is better.
Metric: Correctness (%, human-scored; 330 scenes; correct only with the exact minimum set of objects to remove). Source: arxiv.org. Saturation forecast: Around May 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.5 397B A17B (Thinking) | 35.76 | #242 (Qwen 3.5 397B A17B) |
| 2 | GPT-5.2 (Thinking) | 24.24 | #172 (GPT-5.2) |
| 3 | Qwen 3 VL 30B A3B (Thinking) | 13.94 | #710 |
| 4 | GLM-4.6V (Thinking) | 9.09 | #698 (GLM-4.6V) |
| 5 | Qwen 3 VL 8B (Thinking) | 8.79 | #887 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=3d-scene-structure-multi-object-removal · How It Works · Data refreshed daily, snapshot 2026-11-12.