Spatial Reasoning on VSR
92.61LLM-Judge AccuracyQwen3-VL-235B
Evaluation results
| Method | Links | |
|---|---|---|
| Qwen3-VL-235BToken Retention Ratio=100%2026.05 | 92.61 | |
| F3AToken Retention Ratio=60%2026.05 | 92.39 | |
| VisionZipToken Retention Ratio=60%2026.05 | 92.33 | |
| DivPruneToken Retention Ratio=60%2026.05 | 92.18 | |
| FastVToken Retention Ratio=60%2026.05 | 92.11 | |
| CDPrunerToken Retention Ratio=60%2026.05 | 92.05 | |
| F3AToken Retention Ratio=40%2026.05 | 91.41 | |
| VisionZipToken Retention Ratio=40%2026.05 | 90.27 | |
| CDPrunerToken Retention Ratio=40%2026.05 | 90.02 | |
| FastVToken Retention Ratio=40%2026.05 | 89.98 | |
| DivPruneToken Retention Ratio=40%2026.05 | 89.12 | |
| DivPruneToken Retention Ratio=20%2026.05 | 88.22 | |
| CDPrunerToken Retention Ratio=20%2026.05 | 87.94 | |
| VisionZipToken Retention Ratio=20%2026.05 | 87.9 | |
| F3AToken Retention Ratio=20%2026.05 | 87.81 | |
| FastVToken Retention Ratio=20%2026.05 | 86.31 | |
| PixelReasoner-7B + OursBackbone=PixelReasoner-7B, Modulated thinking process=True2026.03 | 84.3 | |
| PixelReasoner-7BBackbone=PixelReasoner-7B2026.03 | 82.4 | |
| GRIT-3B + OursBackbone=GRIT-3B, Modulated thinking process=True2026.03 | 76.1 | |
| R1-Onevision-7B + OursBackbone=R1-Onevision-7B, Modulated thinking process=True2026.03 | 74.9 | |
| R1-Onevision-7BBackbone=R1-Onevision-7B2026.03 | 73.4 | |
| GRIT-3BBackbone=GRIT-3B2026.03 | 72.9 | |
| Llama11BInter Head=after2026.03 | 66.91 | |
| Llama11BInter Head=before2026.03 | 65.82 | |
| Intern8BInter Head=after2026.03 | 65.28 | |
| Intern8BInter Head=before2026.03 | 64.35 | |
| Qwen7BInter Head=after2026.03 | 63.72 | |
| Qwen7BInter Head=before2026.03 | 62.47 |