Fine-grained 4D Spatial Reasoning on DSR-Bench (test)
87Absolute Distance Accuracy (A.Dis)DSR Suite-Model
Evaluation results
| Method | Links | |||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| DSR Suite-ModelCategory=Spatial Understanding Models2026.05 | 87 | 73.8 | 84.1 | 73.8 | 72 | 35.5 | 75.8 | 76.1 | 77.7 | 60.6 | 37.1 | 35.1 | 46.4 | 58.9 | — | |
| 4DThinker (+DIFT+4DRL)Backbone=InternVL3.5-38B, Training Protocol=+DIFT+4DRL2026.05 | 85.5 | 73.4 | 83.5 | 80.2 | 60.4 | 37.4 | 67.9 | 74.7 | 70.5 | 64.9 | 44.9 | 37.3 | 49.8 | 59.4 | — | |
| 4DThinker (+DIFT+4DRL)Backbone=Qwen3-VL-32B, Training Protocol=+DIFT+4DRL2026.05 | 84.3 | 79.7 | 82.3 | 84 | 67 | 47.8 | 67 | 82.8 | 79.5 | 62.9 | 44.2 | 41 | 48.5 | 62 | — | |
| 4DThinker (+DIFT+4DRL)Backbone=Qwen3-VL-8B, Training Protocol=+DIFT+4DRL2026.05 | 78.3 | 65.6 | 82.3 | 59.3 | 58.2 | 46.1 | 59.4 | 63.2 | 61.5 | 58.8 | 39.1 | 36.6 | 47.2 | 54.6 | — | |
| 4DThinker (+DIFT+4DRL)Backbone=Qwen2.5-VL-7B, Training Protocol=+DIFT+4DRL2026.05 | 71.1 | 62.5 | 60.8 | 70.4 | 49.5 | 39.1 | 60.4 | 47.1 | 43.6 | 38.1 | 32.6 | 32.3 | 42.5 | 47.1 | — | |
| 4DThinker (+DIFT+4DRL)Backbone=InternVL3.5-8B, Training Protocol=+DIFT+4DRL2026.05 | 69.9 | 64.1 | 65.8 | 53.1 | 51.6 | 39.1 | 52.8 | 55.2 | 52.6 | 47.4 | 39.1 | 36.6 | 49.8 | 50 | — | |
| 4DThinker (+DIFT)Backbone=InternVL3.5-38B, Training Protocol=+DIFT2026.05 | 59 | 53.1 | 49.4 | 40.7 | 41.8 | 32.2 | 51.9 | 41.4 | 38.5 | 44.3 | 39.9 | 29.8 | 44.2 | 42.5 | — | |
| VG-LLMCategory=Spatial Understanding Models2026.05 | 55.2 | 32.3 | 58.5 | 57.1 | 51.6 | 32.2 | 56 | 36.3 | 32 | 30.3 | 32.1 | 29.1 | 27.9 | 38.4 | — | |
| 4DThinker (+DIFT)Backbone=Qwen3-VL-32B, Training Protocol=+DIFT2026.05 | 53 | 65.6 | 40.5 | 45.7 | 56 | 41.7 | 50.9 | 48.3 | 46.2 | 43.3 | 37.7 | 34.8 | 44.2 | 45.2 | — | |
| 4DThinker (+DIFT)Backbone=Qwen3-VL-8B, Training Protocol=+DIFT2026.05 | 51.8 | 45.3 | 55.7 | 44.4 | 37.4 | 42.6 | 50 | 47.1 | 44.9 | 42.3 | 35.5 | 32.9 | 40.8 | 42.6 | — | |
| 4DThinker (+DIFT+4DRL)Backbone=Qwen2.5-VL-3B, Training Protocol=+DIFT+4DRL2026.05 | 50.6 | 34.4 | 29.1 | 56.8 | 34.1 | 34.8 | 49.1 | 33.3 | 30.8 | 27.8 | 26.8 | 28.6 | 27.5 | 34.2 | — | |
| 4DThinker (+DIFT)Backbone=Qwen2.5-VL-3B, Training Protocol=+DIFT2026.05 | 48.2 | 28.1 | 25.3 | 46.9 | 28.6 | 28.7 | 35.8 | 29.9 | 37.2 | 24.7 | 25.3 | 31.1 | 26.6 | 31.1 | — | |
| 4DThinker (+DIFT)Backbone=Qwen2.5-VL-7B, Training Protocol=+DIFT2026.05 | 45.8 | 35.9 | 34.2 | 50.6 | 37.4 | 29.6 | 40.6 | 37.9 | 35.9 | 30.9 | 19.6 | 27.3 | 38.6 | 34.8 | — | |
| 4DThinker (+DIFT)Backbone=InternVL3.5-8B, Training Protocol=+DIFT2026.05 | 43.4 | 42.2 | 36.7 | 37 | 33 | 30.4 | 41.5 | 39.1 | 37.2 | 35.1 | 30.4 | 28 | 41.2 | 36.2 | — | |
| VLM-3RCategory=Spatial Understanding Models2026.05 | 28.2 | 27.6 | 31.7 | 42.8 | 38.7 | 33 | 34.4 | 23.8 | 30.8 | 22.2 | 26.4 | 29.1 | 35 | 31.4 | — | |
| InternVL3.5-38BBackbone=InternVL3.5-38B2026.05 | 25.8 | 27.8 | 29.2 | 34.2 | 24.7 | 28.5 | 26.7 | 16.3 | 23.4 | 29.2 | 32.1 | 15.4 | 31.3 | 26.7 | — | |
| Qwen3-VL-8BBackbone=Qwen3-VL-8B2026.05 | 23.5 | 24.6 | 42.6 | 29.7 | 27.9 | 33.8 | 18.1 | 28.4 | 34.5 | 24.2 | 22.1 | 27.9 | 33.5 | 28.7 | — | |
| InternVL3.5-8BBackbone=InternVL3.5-8B2026.05 | 23.5 | 27.6 | 28 | 34.5 | 24.7 | 27.9 | 22.4 | 17 | 19.7 | 28.2 | 30 | 14.2 | 30.1 | 25.4 | — | |
| GPT-5Category=Proprietary Models2026.05 | 21.1 | 41.5 | 48.7 | 34.5 | 33.3 | 34.7 | 17.2 | 44.3 | 41.9 | 21.2 | 25 | 30.9 | 26.7 | 30.8 | — | |
| Qwen2.5-VL-3BBackbone=Qwen2.5-VL-3B2026.05 | 20.5 | 20.3 | 20.3 | 37 | 17.6 | 22.6 | 25.5 | 20.7 | 25.6 | 23.7 | 29 | 24.2 | 27 | 24.6 | — | |
| Gemini-2.5-ProCategory=Proprietary Models2026.05 | 20 | 44.6 | 53.6 | 27.3 | 38.7 | 30.5 | 23.2 | 32.9 | 43.2 | 17.1 | 28.5 | 27.9 | 34.3 | 31.7 | — | |
| Qwen2.5-VL-7BBackbone=Qwen2.5-VL-7B2026.05 | 18.8 | 15.3 | 14.6 | 42.8 | 29 | 19.4 | 31.8 | 19.3 | 11.1 | 22.2 | 19.2 | 20.2 | 30.1 | 23.5 | — | |
| Qwen3-VL-32BBackbone=Qwen3-VL-32B2026.05 | 14.5 | 37.5 | 24.1 | 43.2 | 27.5 | 38.3 | 13.2 | 32.2 | 38.5 | 17.5 | 20.3 | 28 | 31.8 | 28 | — | |
| InternVL2.5Scale=2B2026.05 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 32.5 | |
| InternVL2.5 + DRScaffoldScale=2B2026.05 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 66.6 | |
| Phi4-multimodalScale=5.6B2026.05 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 45.8 | |
| Phi4-multimodal + DRScaffoldScale=5.6B2026.05 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 50.7 | |
| Qwen2.5-VLScale=3B2026.05 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 35 | |
| Qwen2.5-VLScale=32B2026.05 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 44.7 | |
| Qwen2.5-VL + DRScaffoldScale=3B2026.05 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 62.6 |