3D Scene Structure - Single-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): shown a synthetic photorealistic indoor scene, the model names every object that would become visible if a given object were removed, which requires inferring depth-ordered occlusion. Trained annotators grade every response with no partial credit; models run with thinking enabled at temperature 0. Correctness in percent over 602 scenes, credited only for the complete set with nothing invented; higher is better.
Metric: Correctness (%, human-scored; 602 scenes; correct only with complete recall of the newly visible objects and nothing invented). Source: arxiv.org. Saturation forecast: Around 2029. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.2 (Thinking) | 23.01 | #172 (GPT-5.2) |
| 2 | Qwen 3.5 397B A17B (Thinking) | 22.85 | #242 (Qwen 3.5 397B A17B) |
| 3 | Qwen 3 VL 30B A3B (Thinking) | 21.03 | #710 |
| 4 | Qwen 3 VL 8B (Thinking) | 14.9 | #887 |
| 5 | GLM-4.6V (Thinking) | 14.57 | #698 (GLM-4.6V) |
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-single-object-removal · How It Works · Data refreshed daily, snapshot 2026-11-12.