Procedural assembly reasoning on LEGO-Puzzles
52Height AccuracyInternVL3-78B
Evaluation Results
| Method | Links | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| InternVL3-78BAccess=Open-source2026.02 | 52 | 63 | 36 | 54 | 64 | 80 | 58 | 29 | 25 | 22 | 37 | 47.3 | |
| Gemini-2.5-FlashAccess=Proprietary2026.02 | 52 | 58 | 37 | 55 | 58 | 74 | 53 | 49 | 40 | 46 | 29 | 50.1 | |
| GPT-4oAccess=Proprietary2026.02 | 49 | 66 | 41 | 51 | 65 | 87 | 51 | 51 | 53 | 72 | 49 | 57.7 | |
| Qwen2.5-VL-72BAccess=Open-source2026.02 | 43 | 58 | 38 | 39 | 57 | 76 | 57 | 52 | 74 | 43 | 43 | 52.7 | |
| InternVL3-38BAccess=Open-source2026.02 | 40 | 60 | 39 | 55 | 57 | 81 | 59 | 33 | 47 | 12 | 36 | 47.2 | |
| Claude-3.5-SonnetAccess=Proprietary2026.02 | 39 | 60 | 42 | 48 | 61 | 78 | 58 | 37 | 49 | 54 | 64 | 53.6 | |
| Qwen2.5-VL-32BAccess=Open-source2026.02 | 35 | 60 | 38 | 52 | 45 | 79 | 51 | 45 | 66 | 43 | 43 | 50.6 | |
| TPRU-32BAccess=Ours, Backbone=Qwen2.5-VL-32B2026.02 | 34 | 61 | 35 | 47 | 55 | 76 | 52 | 48 | 70 | 49 | 48 | 52.3 | |
| TPRU-3BAccess=Ours, Backbone=Qwen2.5-VL-3B2026.02 | 30 | 54 | 32 | 40 | 37 | 69 | 49 | 14 | 30 | 8 | 24 | 35.2 | |
| Qwen2.5-VL-3BAccess=Open-source2026.02 | 29 | 55 | 30 | 36 | 32 | 65 | 48 | 19 | 16 | 4 | 25 | 32.6 | |
| TPRU-7BAccess=Ours, Backbone=Qwen2.5-VL-7B2026.02 | 23 | 56 | 37 | 40 | 45 | 67 | 55 | 31 | 49 | 32 | 36 | 42.8 | |
| Qwen2.5-VL-7BAccess=Open-source2026.02 | 20 | 57 | 32 | 47 | 38 | 67 | 56 | 29 | 22 | 8 | 25 | 36.5 |