Spatio-temporal reasoning on Dyn-Bench
84.5Scene Focus & Dyn. ScoreQwen3-VL-8B + DIFT + 4DRL
Evaluation results
| Method | Links | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Qwen3-VL-8B + DIFT + 4DRLBase Model=Qwen3-VL-8B, Training Strategy=DIFT+4DRL2026.05 | 84.5 | — | 74.5 | — | — | — | — | — | — | 68 | — | 70.9 | |
| Qwen3-VL-235Btype=Open-source Model, parameters=235B2026.03 | 84.1 | 1 | 65.3 | 76.4 | 55.8 | 55.6 | 77.8 | 76.1 | 59.8 | 59 | 60.2 | — | |
| Qwen3-VL-32B + DIFT + 4DRLBase Model=Qwen3-VL-32B, Training Strategy=DIFT+4DRL2026.05 | 84 | — | 75.4 | — | — | — | — | — | — | 68.7 | — | 73.6 | |
| Qwen3-VL-8Btype=Open-source Model, parameters=8B2026.03 | 82.4 | 3 | 61.4 | 70.8 | 52.6 | 53.6 | 75 | 71.2 | 55.4 | 52.6 | 60 | — | |
| Qwen3-VL-32Btype=Open-source Model, parameters=32B2026.03 | 80.2 | 2 | 62.7 | 73.7 | 56.2 | 53.4 | 74.6 | 73.1 | 58.2 | 54.3 | 56.9 | — | |
| Qwen3-VL-32B + DIFTBase Model=Qwen3-VL-32B, Training Strategy=DIFT2026.05 | 79.8 | — | 69.2 | — | — | — | — | — | — | 61.5 | — | 66.3 | |
| UniPixel-7Btype=Region-level MLLM, parameters=7B2026.03 | 79.7 | 1 | 58.1 | 64.4 | 50.2 | 54.7 | 76.1 | 70.4 | 47.3 | 47.3 | 55.7 | — | |
| VST-7B-RLtype=Spatial MLLM, parameters=7B2026.03 | 79.4 | 2 | 55.7 | 68.6 | 48.4 | 51.9 | 73 | 70.7 | 45.1 | 39.1 | 52.9 | — | |
| Qwen2.5-VL-7B + DIFT + 4DRLBase Model=Qwen2.5-VL-7B, Training Strategy=DIFT+4DRL2026.05 | 79.2 | — | 65.9 | — | — | — | — | — | — | 54.9 | — | 63.5 | |
| Qwen3-VL-8B + DIFTBase Model=Qwen3-VL-8B, Training Strategy=DIFT2026.05 | 79 | — | 68 | — | — | — | — | — | — | 60.8 | — | 64.1 | |
| SpaceR-7Btype=Spatial MLLM, parameters=7B2026.03 | 78.2 | 1 | 56.5 | 66.6 | 49.2 | 52.7 | 72.2 | 67.8 | 50.3 | 40 | 55.5 | — | |
| InternVL3-14Btype=Open-source Model, parameters=14B2026.03 | 77.7 | 7 | 53.7 | 65.3 | 47 | 49.7 | 67.8 | 69.2 | 37.9 | 44.9 | 46.9 | — | |
| UniPixel-3Btype=Region-level MLLM, parameters=3B2026.03 | 77.7 | 2 | 55.4 | 63.3 | 47.2 | 53.2 | 71.7 | 70.2 | 43.2 | 43.6 | 52 | — | |
| SpatialLadder-3Btype=Spatial MLLM, parameters=3B2026.03 | 77.1 | 3 | 53.6 | 60.8 | 46.1 | 49.2 | 70 | 70.9 | 38.2 | 42 | 51.9 | — | |
| InternVL3-38Btype=Open-source Model, parameters=38B2026.03 | 76.7 | 5 | 54.2 | 68.2 | 44.8 | 48.4 | 71.1 | 67.3 | 41.2 | 48.9 | 44.6 | — | |
| InternVL3.5-38B + DIFT + 4DRLBase Model=InternVL3.5-38B, Training Strategy=DIFT+4DRL2026.05 | 76.6 | — | 65.8 | — | — | — | — | — | — | 57.9 | — | 62.9 | |
| Qwen3-VL-8BModel Category=Different Base Models (4DThinker)2026.05 | 76.2 | — | 63.7 | — | — | — | — | — | — | 56 | — | 59 | |
| Qwen3-VL-32BModel Category=Different Base Models (4DThinker)2026.05 | 76 | — | 64.5 | — | — | — | — | — | — | 56.5 | — | 61.1 | |
| InternVL3.5-8B + DIFT + 4DRLBase Model=InternVL3.5-8B, Training Strategy=DIFT+4DRL2026.05 | 75.5 | — | 64.5 | — | — | — | — | — | — | 54.9 | — | 63.1 | |
| LLaVA-OneVision-1.5-8Btype=Open-source Model, parameters=8B, version=1.52026.03 | 75.4 | 6 | 53.8 | 60.9 | 47.7 | 53.4 | 74.4 | 69.6 | 41 | 37 | 51.6 | — | |
| Qwen2.5-VL-32Btype=Open-source Model, parameters=32B2026.03 | 75.3 | 4 | 56 | 71.5 | 52.2 | 53.8 | 71.5 | 67.6 | 42.1 | 46.7 | 47.9 | — | |
| Sa2VA-InternVL3-14Btype=Region-level MLLM, parameters=14B, backbone=InternVL32026.03 | 74.6 | 3 | 53.6 | 55.9 | 48.9 | 53.2 | 72 | 70.2 | 38.1 | 39.6 | 53.6 | — | |
| VST-7B-RLModel Category=Spatial Understanding Models2026.05 | 74.4 | — | 58.8 | — | — | — | — | — | — | 45.7 | — | 56.3 | |
| Qwen2.5-VL-7B + DIFTBase Model=Qwen2.5-VL-7B, Training Strategy=DIFT2026.05 | 74.4 | — | 59.5 | — | — | — | — | — | — | 47.8 | — | 56.3 | |
| Spatial-SSRL-7Btype=Spatial MLLM, parameters=7B2026.03 | 73.8 | 4 | 45.9 | 54.5 | 40 | 48.1 | 68.5 | 65.9 | 35.8 | 36.7 | 37.7 | — | |
| Qwen2.5-VL-7Btype=Open-source Model, parameters=7B2026.03 | 73.7 | 9 | 51.6 | 61.7 | 42.8 | 48 | 69.1 | 67 | 43.8 | 39.6 | 42.8 | — | |
| Sa2VA-Qwen3-VL-4Btype=Region-level MLLM, parameters=4B, backbone=Qwen3-VL2026.03 | 73.2 | 5 | 49.8 | 60.8 | 39.3 | 46.2 | 67.2 | 62 | 41 | 44.5 | 36.8 | — | |
| GPT-5type=Proprietary Model (API)2026.03 | 73 | 2 | 59.5 | 68.6 | 47.3 | 48.1 | 71.7 | 65.9 | 60.9 | 58.4 | 58.4 | — | |
| SpaceR-7BModel Category=Spatial Understanding Models2026.05 | 72.7 | — | 59.2 | — | — | — | — | — | — | 48.6 | — | 56.2 | |
| SpatialLadder-3BModel Category=Spatial Understanding Models2026.05 | 72.7 | — | 56.2 | — | — | — | — | — | — | 44 | — | 52 | |
| InternVL3.5-38Btype=Open-source Model, parameters=38B2026.03 | 72.3 | 10 | 50.8 | 65 | 40.7 | 42.4 | 60.2 | 62.3 | 45.5 | 42.9 | 44.6 | — | |
| Sa2VA-InternVL2.5-8Btype=Region-level MLLM, parameters=8B, backbone=InternVL2.52026.03 | 71.9 | 6 | 49.4 | 61 | 42.4 | 45.7 | 66.1 | 62.8 | 36.6 | 36.4 | 47.2 | — | |
| Sa2VA-Qwen2.5-VL-7Btype=Region-level MLLM, parameters=7B, backbone=Qwen2.5-VL2026.03 | 70.5 | 4 | 50.3 | 58.6 | 39.3 | 52.9 | 67.6 | 62.1 | 38.8 | 39.1 | 49.3 | — | |
| GPT-5Model Category=Proprietary Models2026.05 | 70.2 | — | 61.4 | — | — | — | — | — | — | 59.2 | — | 54.7 | |
| InternVL3.5-38B + DIFTBase Model=InternVL3.5-38B, Training Strategy=DIFT2026.05 | 70.1 | — | 58.6 | — | — | — | — | — | — | 50.3 | — | 55.4 | |
| Qwen2.5-VL-7BModel Category=Different Base Models (4DThinker)2026.05 | 69.9 | — | 54.3 | — | — | — | — | — | — | 42.1 | — | 50.8 | |
| InternVL3.5-8B + DIFTBase Model=InternVL3.5-8B, Training Strategy=DIFT2026.05 | 69.8 | — | 58.1 | — | — | — | — | — | — | 48.2 | — | 56.3 | |
| InternVL3.5-8Btype=Open-source Model, parameters=8B2026.03 | 69.6 | 11 | 50.3 | 66.6 | 41 | 44.8 | 63.5 | 62.7 | 41.5 | 40.1 | 46.4 | — | |
| Spatial-SSRL-7BModel Category=Spatial Understanding Models2026.05 | 69.4 | — | 51.2 | — | — | — | — | — | — | 36.7 | — | 47.5 | |
| GPT-4otype=Proprietary Model (API)2026.03 | 68.8 | 3 | 50.1 | 56.1 | 38.7 | 44.6 | 63.1 | 59.1 | 47.2 | 42 | 49.2 | — | |
| Qwen2.5-VL-72Btype=Open-source Model, parameters=72B2026.03 | 68.3 | 8 | 51.8 | 65.5 | 41.1 | 43.7 | 60.4 | 57.8 | 49.2 | 42.6 | 55 | — | |
| Gemini-2.5 Protype=Proprietary Model (API)2026.03 | 65.6 | 1 | 59.8 | 69.7 | 48 | 50.5 | 67.8 | 59.9 | 60.7 | 54.9 | 51.8 | — | |
| InternVL3.5-8BModel Category=Different Base Models (4DThinker)2026.05 | 65.3 | — | 52.9 | — | — | — | — | — | — | 42.7 | — | 50.8 | |
| InternVL3.5-38BModel Category=Different Base Models (4DThinker)2026.05 | 64.9 | — | 52.9 | — | — | — | — | — | — | 44.3 | — | 49.4 | |
| Gemini-2.5-ProModel Category=Proprietary Models2026.05 | 64.4 | — | 58.8 | — | — | — | — | — | — | 55.8 | — | 56.1 | |
| LLaVA-OneVision-1.5-4Btype=Open-source Model, parameters=4B, version=1.52026.03 | 63.9 | 12 | 49.9 | 50.5 | 48.5 | 50.3 | 65.2 | 64.8 | 39.9 | 36.6 | 46.1 | — | |
| VideoGLaMMtype=Region-level MLLM2026.03 | 39.3 | 7 | 30.7 | 35.6 | 34.4 | 35 | 34.6 | 38.2 | 22.7 | 21.2 | 25.9 | — | |
| Chance Level (Random)type=Baseline2026.03 | 25 | — | — | 25 | 25 | 25 | 25 | 25 | 25 | 25 | 25 | — | |
| Chance Level (Frequency)type=Baseline2026.03 | 13.8 | — | — | 12.3 | 21.6 | 31.5 | 29.1 | 27.7 | 33.5 | 10.1 | 25.6 | — |