IDEAL-Bench - Geometric Reasoning: leaderboard
Metric: Geometric score (0-100): unweighted mean of position accuracy, box precision and rotation accuracy (non-grid rooms) and grid coverage and grid self-consistency (classroom, library and office scenes) on the IDEAL-Bench 3D layout task; higher is better. Source: arxiv.org. Saturation forecast: Around 2030. 13 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 2.5 Pro | 40.1 |
| 2 | GPT-4o | 35.9 |
| 3 | Gemma 4 31B (IT) | 35.7 |
| 4 | Claude Sonnet 4.6 | 34.7 |
| 5 | GLM-4.6V | 30.7 |
| 6 | Qwen 3 VL 8B Instruct | 28.6 |
| 7 | GPT-5.4 | 28.4 |
| 8 | Qwen 2.5 VL 72B Instruct | 27.9 |
| 9 | InternVL3.5-8B | 27 |
| 10 | Qwen 3 VL 235B A22B Instruct | 24.3 |
| 11 | Qwen 3 VL 30B A3B Instruct | 22.6 |
Interactive version: theaggregate.ai/benchmark?slug=ideal-bench-geometric-reasoning · How It Works · Data refreshed daily, snapshot 2026-09-29.