Visual token compression evaluation on VTCBench
57.2GUQMoP
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| QMoPType=MoE2026.03 | 57.2 | 48.6 | 64.7 | 75 | 60.6 | 61.8 | |
| LDP-V2Type=Pooling (B)2026.03 | 56.8 | 48.3 | 63.9 | 74.3 | 56.4 | 60.5 | |
| LLaVA-1.5-7BType=Full-size (Baseline)2026.03 | 56.1 | 47.4 | 65.9 | 72 | 60.9 | 61 | |
| TokenPackerType=Pooling (B)2026.03 | 55.8 | 47.8 | 64.7 | 74 | 57.5 | 60.5 | |
| MQT-LLaVAType=Resampler (A)2026.03 | 54.7 | 46.1 | 64.3 | 74.7 | 54.9 | 59.5 | |
| FasterVLMType=Pruning (C)2026.03 | 54.7 | 43.5 | 59.6 | 70.7 | 58.3 | 58 | |
| Pixel-ShuffleType=Pooling (B)2026.03 | 53.6 | 47 | 61.2 | 70 | 55.3 | 57.9 | |
| C-AbstractType=Pooling (B)2026.03 | 53.2 | 47.8 | 59.6 | 70.3 | 55.2 | 57.7 |