Visual Spatial Instruction on VSI-Bench
40.5Relative Direction (Hard)LLaVA-OneVision-1.5-4B + SpatialSV
Evaluation Results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| LLaVA-OneVision-1.5-4B + SpatialSVSupervision=SpatialSV, Model Family=LLaVA-OneVision-1.5, Model Size=4B2026.06 | 40.5 | 35.2 | 52.5 | 42 | 25.4 | 36.1 | 37.1 | |
| LLaVA-OneVision-1.5-4B + DistillationSupervision=Feature Distillation, Model Family=LLaVA-OneVision-1.5, Model Size=4B2026.06 | 40 | 35.5 | 47.9 | 39 | 24.6 | 33 | 35.3 | |
| LLaVA-OneVision-1.5-4B + TextSupervision=Textual, Model Family=LLaVA-OneVision-1.5, Model Size=4B2026.06 | 38.9 | 34.1 | 48.9 | 38.6 | 25.1 | 30.9 | 34.9 | |
| LLaVA-OneVision-1.5-8B + SpatialSVSupervision=SpatialSV, Model Family=LLaVA-OneVision-1.5, Model Size=8B2026.06 | 38.3 | 36.5 | 49.3 | 42.1 | 29.6 | 40.2 | 38.1 | |
| LLaVA-OneVision-1.5-4BSupervision=None, Model Family=LLaVA-OneVision-1.5, Model Size=4B2026.06 | 37.3 | 35.5 | 46.1 | 37 | 24.3 | 29.9 | 33.9 | |
| LLaVA-OneVision-1.5-8B + DistillationSupervision=Feature Distillation, Model Family=LLaVA-OneVision-1.5, Model Size=8B2026.06 | 36.5 | 35.5 | 50.2 | 38.3 | 28.3 | 32 | 35.7 | |
| Qwen2.5-VL-3B + DistillationSupervision=Feature Distillation, Model Family=Qwen2.5-VL, Model Size=3B2026.06 | 35.9 | 34.1 | 30.9 | 32.1 | 22.8 | 32 | 30.6 | |
| Qwen2.5-VL-3B + TextSupervision=Textual, Model Family=Qwen2.5-VL, Model Size=3B2026.06 | 35.7 | 35.7 | 26.3 | 31.1 | 23.6 | 29.9 | 30.1 | |
| Qwen2.5-VL-3B + SpatialSVSupervision=SpatialSV, Model Family=Qwen2.5-VL, Model Size=3B2026.06 | 34.6 | 37.8 | 34.6 | 33.5 | 23.3 | 37.1 | 32.2 | |
| Qwen2.5-VL-3BSupervision=None, Model Family=Qwen2.5-VL, Model Size=3B2026.06 | 34.3 | 33.3 | 29 | 31.7 | 22.7 | 27.8 | 29.6 | |
| LLaVA-OneVision-1.5-8B + TextSupervision=Textual, Model Family=LLaVA-OneVision-1.5, Model Size=8B2026.06 | 34.1 | 35.5 | 52.1 | 37.3 | 29.9 | 30.4 | 35.5 | |
| InternVL3-2B + SpatialSVSupervision=SpatialSV, Model Family=InternVL3, Model Size=2B2026.06 | 31.4 | 35.2 | 47.9 | 35.1 | 28.8 | 41.2 | 34.6 | |
| InternVL3-1BSupervision=None, Model Family=InternVL3, Model Size=1B2026.06 | 30.8 | 33.9 | 47 | 26.6 | 24.6 | 33.5 | 30.2 | |
| InternVL3-8B + SpatialSVSupervision=SpatialSV, Model Family=InternVL3, Model Size=8B2026.06 | 30.6 | 37.3 | 50.7 | 36.2 | 33.2 | 42.8 | 36.6 | |
| InternVL3-1B + SpatialSVSupervision=SpatialSV, Model Family=InternVL3, Model Size=1B2026.06 | 30 | 34.7 | 50.2 | 28.2 | 24.9 | 38.7 | 31.4 | |
| InternVL3-1B + DistillationSupervision=Feature Distillation, Model Family=InternVL3, Model Size=1B2026.06 | 29.8 | 35.2 | 43.8 | 28.2 | 25.1 | 32 | 30.4 | |
| Qwen2.5-VL-7B + SpatialSVSupervision=SpatialSV, Model Family=Qwen2.5-VL, Model Size=7B2026.06 | 29.5 | 35.2 | 48.4 | 37.2 | 30.1 | 42.8 | 35.4 | |
| LLaVA-NeXT-Video-7B + DistillationSupervision=Feature Distillation, Model Family=LLaVA-NeXT-Video, Model Size=7B2026.06 | 29.5 | 33.6 | 48.4 | 37.5 | 29.5 | 34.5 | 34.4 | |
| InternVL3-2B + DistillationSupervision=Feature Distillation, Model Family=InternVL3, Model Size=2B2026.06 | 29.5 | 33.3 | 47 | 33 | 28 | 31.4 | 32.4 | |
| LLaVA-NeXT-Video-7B + SpatialSVSupervision=SpatialSV, Model Family=LLaVA-NeXT-Video, Model Size=7B2026.06 | 29.2 | 33.9 | 52.5 | 37.8 | 30.4 | 44.9 | 35.9 | |
| Qwen2.5-VL-7B + DistillationSupervision=Feature Distillation, Model Family=Qwen2.5-VL, Model Size=7B2026.06 | 29 | 32.5 | 50.2 | 34.8 | 29.1 | 37.6 | 33.7 | |
| InternVL3-1B + TextSupervision=Textual, Model Family=InternVL3, Model Size=1B2026.06 | 29 | 34.7 | 44.7 | 27.3 | 25.4 | 30.9 | 30 | |
| InternVL3-2B + TextSupervision=Textual, Model Family=InternVL3, Model Size=2B2026.06 | 29 | 34.1 | 45.6 | 32 | 30.1 | 29.9 | 32.4 | |
| LLaVA-OneVision-1.5-8BSupervision=None, Model Family=LLaVA-OneVision-1.5, Model Size=8B2026.06 | 28.7 | 34.9 | 48.9 | 36.1 | 28.8 | 28.9 | 33.5 | |
| Qwen2.5-VL-7BSupervision=None, Model Family=Qwen2.5-VL, Model Size=7B2026.06 | 28.4 | 24.9 | 45.2 | 31.8 | 29.5 | 31.4 | 30.8 | |
| InternVL3-8B + DistillationSupervision=Feature Distillation, Model Family=InternVL3, Model Size=8B2026.06 | 28.4 | 36 | 47.5 | 35.5 | 31.7 | 34 | 34.5 | |
| LLaVA-NeXT-Video-7B + TextSupervision=Textual, Model Family=LLaVA-NeXT-Video, Model Size=7B2026.06 | 27.9 | 35.7 | 45.6 | 35.5 | 29.8 | 32.5 | 33.6 | |
| Qwen2.5-VL-7B + TextSupervision=Textual, Model Family=Qwen2.5-VL, Model Size=7B2026.06 | 27.4 | 31.2 | 51.2 | 31.4 | 30.1 | 34.5 | 32.4 | |
| LLaVA-NeXT-Video-7BSupervision=None, Model Family=LLaVA-NeXT-Video, Model Size=7B2026.06 | 27.1 | 32.5 | 50.7 | 35.9 | 27.5 | 31.4 | 32.9 | |
| Chance Level (Random)Type=Baseline2026.06 | 26.5 | 25.7 | 24.9 | 24.1 | 23.3 | 28.4 | 27.7 | |
| InternVL3-2BSupervision=None, Model Family=InternVL3, Model Size=2B2026.06 | 26.5 | 31.5 | 48.4 | 30.1 | 24.4 | 34 | 30.3 | |
| InternVL3-8B + TextSupervision=Textual, Model Family=InternVL3, Model Size=8B2026.06 | 25.7 | 33.6 | 50.2 | 34.5 | 32 | 29.9 | 33.5 | |
| Chance Level (Frequency)Type=Baseline2026.06 | 25.2 | 33.6 | 50.2 | 25.1 | 25.2 | 29.4 | 29 | |
| InternVL3-8BSupervision=None, Model Family=InternVL3, Model Size=8B2026.06 | 23.6 | 34.1 | 46.1 | 34.1 | 32.9 | 32.5 | 33.1 |