Spatial Alignment on RoomEditing
98.3mIoU3D-Layout-R1 (Qwen3-8B)
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| 3D-Layout-R1 (Qwen3-8B)Backbone=Qwen3-8B, Training strategy=Ours2026.03 | 98.3 | 98.3 | 0.031 | |
| 3D-Layout-R1 (Qwen2.5-7B)Backbone=Qwen2.5-7B, Training strategy=Ours2026.03 | 97.4 | 97.3 | 0.033 | |
| Qwen3-8B (CoT-SFT)Backbone=Qwen3-8B, Training strategy=CoT-SFT2026.03 | 97.2 | 97.2 | 0.026 | |
| Qwen2.5-7B (CoT-SFT)Backbone=Qwen2.5-7B, Training strategy=CoT-SFT2026.03 | 95 | 95 | 0.051 | |
| Qwen3-8B (Ans-SFT)Backbone=Qwen3-8B, Training strategy=Ans-SFT2026.03 | 89.5 | 89.7 | 0.117 | |
| Qwen2.5-7B (Ans-SFT)Backbone=Qwen2.5-7B, Training strategy=Ans-SFT2026.03 | 86.9 | 86.3 | 0.138 | |
| Qwen3-8B (Vanilla GRPO)Backbone=Qwen3-8B, Training strategy=Vanilla GRPO2026.03 | 76.8 | 77.4 | 0.336 | |
| Gemini 2.5 Pro2026.03 | 59.1 | 58.7 | 0.621 | |
| Qwen3-VL-235B-InstructBackbone=Qwen3-VL-235B, Variant=Instruct2026.03 | 50.2 | 49.7 | 0.719 | |
| Qwen3-VL-235B-ThinkingBackbone=Qwen3-VL-235B, Variant=Thinking2026.03 | 49.6 | 48.7 | 0.684 | |
| Deepseek-V32026.03 | 47.3 | 46.9 | 0.859 | |
| Deepseek-R1-05282026.03 | 47.1 | 46.3 | 0.871 | |
| Deepseek-R12026.03 | 45.4 | 44.9 | 0.754 | |
| Gemini 2.5 Flash2026.03 | 44.3 | 44.1 | 0.924 |