Multimodal Understanding on MMMU, MathVista, MMBEn, MMBCn, MMStar, HallBench, AI2D, and OCRBench
63.9MMMU ScorePost-ViT
Evaluation Results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Post-ViTData Scale=64M, Compression Strategy=post-ViT2026.05 | 63.9 | 76.3 | 87 | 86.4 | 67.9 | 56.5 | 84.7 | 86.7 | 76.2 | |
| Post-ViTData Scale=32M, Compression Strategy=post-ViT2026.05 | 63.6 | 72.7 | 85.5 | 84.9 | 65.9 | 53.6 | 82.5 | 84.8 | 74.2 | |
| intra-ViT early compressionData Scale=32M, Compression Strategy=intra-ViT (4x at layer 6, 4x after ViT)2026.05 | 62.3 | 72 | 84.7 | 85 | 66.2 | 52.8 | 82.4 | 82.7 | 73.5 | |
| intra-ViT early compressionData Scale=64M, Compression Strategy=intra-ViT (4x at layer 6, 4x after ViT)2026.05 | 61.9 | 76.9 | 86.2 | 86.5 | 66.9 | 55.2 | 84.9 | 85.9 | 75.6 | |
| intra-ViT early compressionData Scale=16M, Compression Strategy=intra-ViT (4x at layer 6, 4x after ViT)2026.05 | 61.2 | 71.1 | 84.1 | 83.3 | 65.3 | 54.7 | 81.8 | 83.5 | 73.1 | |
| intra-ViT early compressionData Scale=4M, Compression Strategy=intra-ViT (4x at layer 6, 4x after ViT)2026.05 | 60.3 | 61.7 | 78.6 | 78.4 | 60.4 | 47.7 | 76.6 | 75.3 | 67.4 | |
| intra-ViT early compressionData Scale=8M, Compression Strategy=intra-ViT (4x at layer 6, 4x after ViT)2026.05 | 59.6 | 68.6 | 83.4 | 81.6 | 62.9 | 52 | 80.6 | 76.7 | 70.7 | |
| Post-ViTData Scale=16M, Compression Strategy=post-ViT2026.05 | 59.1 | 71 | 84.9 | 83.5 | 65.5 | 51.5 | 81.2 | 83.2 | 72.5 | |
| Post-ViTData Scale=8M, Compression Strategy=post-ViT2026.05 | 58.6 | 67.3 | 83.7 | 82.3 | 62.9 | 51.2 | 79.8 | 79.1 | 70.6 | |
| Post-ViTData Scale=4M, Compression Strategy=post-ViT2026.05 | 57.9 | 63 | 79.4 | 79.1 | 60.6 | 50.5 | 77.7 | 77.5 | 68.2 |