Visual Question Answering on RealWorld
68.89AccuracyHolo-V
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Holo-VRetain Token Ratio Budget=70%2026.03 | 68.89 | 0.7 | |
| Qwen2.5-VL-7BResolution=Full-resolution baseline2026.03 | 68.8 | 1 | |
| SparseVLMRetain Token Ratio Budget=70%2026.03 | 68.5 | 0.7 | |
| AwaResStrategy=Adaptive resolution escalation2026.03 | 68.5 | 0.43 | |
| SparseVLMRetain Token Ratio Budget=50%2026.03 | 68.4 | 0.5 | |
| Holo-VRetain Token Ratio Budget=50%2026.03 | 67.19 | 0.5 | |
| VisionThinkStrategy=Adaptive resolution escalation2026.03 | 66.6 | 0.55 | |
| VisionZIPRetain Token Ratio Budget=50%2026.03 | 66.01 | 0.5 | |
| Qwen2.5-VL-7BResolution=Low-res images2026.03 | 66 | 0.25 | |
| VisionZIPRetain Token Ratio Budget=70%2026.03 | 64.84 | 0.7 | |
| VanillaToken Reduction Ratio=100% (Upper Bound)2026.06 | 59.35 | — | |
| ALVTSToken Reduction Ratio=89%2026.06 | 51.37 | — | |
| FastVToken Reduction Ratio=89%2026.06 | 49.8 | — | |
| PDropToken Reduction Ratio=89%2026.06 | 44.84 | — |