Compositional Reasoning on MMSI-Bench
54.3MSR ScoreSpatiO
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| SpatiOSetting=Zero-shot, Optimization Samples=150 samples from 3DSRBench+STVQA+CV-Bench2026.04 | 54.3 | 41 | 42.1 | 37 | 43.6 | |
| Llava-4DTraining Protocol=Without LoRA training, Setting=Zero-shot2026.04 | 24.7 | 23.2 | 23.1 | 18 | 23.2 | |
| Qwen-3.0-VL-4BTraining Protocol=Without LoRA training, Setting=Zero-shot2026.04 | 24.3 | 22.2 | 26.9 | 23.3 | 24.1 | |
| SpatialReasonerTraining Protocol=Without LoRA training, Setting=Zero-shot2026.04 | 23.8 | 21.2 | 25.4 | 14.7 | 22.1 | |
| SpatialRGPTTraining Protocol=Without LoRA training, Setting=Zero-shot2026.04 | 22.4 | 15 | 17.4 | 3.2 | 17.3 | |
| Sa2VATraining Protocol=Without LoRA training, Setting=Zero-shot2026.04 | 8.6 | 9.6 | 10.8 | 6 | 8.7 |