Diagram Understanding on AI2D (Efficiency Metrics)
87.14AI2D ScoreBaseline
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| BaselineVision Token Retention Ratio=100%, Saving FLOPs. (%)=02025.11 | 87.14 | 79.16 | 0 | — | |
| FastMMoE†Vision Token Retention Ratio=75%, lv=6, Kv=5, Saving FLOPs. (%)=37.592025.11 | 86.37 | 77.4 | 1.76 | — | |
| FastMMoE*Vision Token Retention Ratio=75%, Saving FLOPs. (%)=21.992025.11 | 86.27 | 77.56 | 1.6 | — | |
| FastVVision Token Retention Ratio=75%, Saving FLOPs. (%)=21.992025.11 | 85.59 | 76.93 | 2.22 | — | |
| FastMMoE†Vision Token Retention Ratio=75%, lv=3, Kv=4, Saving FLOPs. (%)=44.652025.11 | 85.46 | 76.84 | 2.31 | — | |
| FastMMoE*Vision Token Retention Ratio=50%, Saving FLOPs. (%)=44.232025.11 | 84.49 | 76.35 | 2.8 | — | |
| InternVL3Params=8B2026.04 | 83.9 | — | — | — | |
| SparseVLMVision Token Retention Ratio=75%, Saving FLOPs. (%)=21.992025.11 | 83.81 | 70.06 | 9.09 | — | |
| FastMMoE†Vision Token Retention Ratio=50%, lv=6, Kv=5, Saving FLOPs. (%)=55.022025.11 | 83.71 | 75.63 | 3.53 | — | |
| FastMMoE†Vision Token Retention Ratio=50%, lv=3, Kv=3, Saving FLOPs. (%)=64.552025.11 | 83.48 | 74.69 | 4.47 | — | |
| Qwen3-VLParams=8B2026.04 | 83.2 | — | — | — | |
| Qwen2.5-VLParams=7B2026.04 | 82.5 | — | — | — | |
| FastVVision Token Retention Ratio=50%, Saving FLOPs. (%)=44.232025.11 | 82.42 | 74.33 | 4.82 | — | |
| BaselineRetention Ratio (Vision Tokens)=100%2025.11 | 82.38 | 72.68 | 0 | 0 | |
| InternVL3.5Params=8B2026.04 | 82.3 | — | — | — | |
| FastMMoE*Vision Token Retention Ratio=25%, Saving FLOPs. (%)=64.302025.11 | 82.12 | 73.52 | 5.63 | — | |
| FastMMoE*Retention Ratio (Vision Tokens)=75%2025.11 | 82.06 | 72.49 | 0.19 | 21.09 | |
| FastVRetention Ratio (Vision Tokens)=75%2025.11 | 81.99 | 72.18 | 0.5 | — | |
| SparseVLMRetention Ratio (Vision Tokens)=75%2025.11 | 81.57 | 71.72 | 0.96 | — | |
| FastMMoE†Retention Ratio (Vision Tokens)=75%, lv=10, Kv=22025.11 | 81.54 | 72 | 0.68 | 39.15 | |
| FastMMoE†Vision Token Retention Ratio=25%, lv=6, Kv=5, Saving FLOPs. (%)=70.472025.11 | 81.38 | 71.89 | 7.27 | — | |
| FastMMoE*Retention Ratio (Vision Tokens)=50%2025.11 | 81.35 | 71.94 | 0.74 | 43.07 | |
| FastVRetention Ratio (Vision Tokens)=50%2025.11 | 81.19 | 71.91 | 0.77 | — | |
| SparseVLMRetention Ratio (Vision Tokens)=50%2025.11 | 80.51 | 70.31 | 2.37 | — | |
| FastMMoE†Retention Ratio (Vision Tokens)=50%, lv=2, Kv=22025.11 | 80.38 | 70.57 | 2.11 | 61.76 | |
| MUSE-3BType=Unified, # Params=2B+1.6B2026.05 | 80.2 | — | — | — | |
| FastMMoE*Retention Ratio (Vision Tokens)=25%2025.11 | 79.15 | 70.32 | 2.36 | 63.66 | |
| SparseVLMRetention Ratio (Vision Tokens)=25%2025.11 | 79.02 | 66.61 | 6.07 | — | |
| FastVRetention Ratio (Vision Tokens)=25%2025.11 | 78.82 | 70.02 | 2.65 | — | |
| Show-o2Type=Unified, # Params=7B2026.05 | 78.6 | — | — | — | |
| UniLIP-3BType=Unified, # Params=2B+1.6B2026.05 | 78.6 | — | — | — | |
| InternVL3Type=Und. Only, # Params=1.8B2026.05 | 78.5 | — | — | — | |
| SparseVLMVision Token Retention Ratio=50%, Saving FLOPs. (%)=44.232025.11 | 77.36 | 62.55 | 16.6 | — | |
| FastVVision Token Retention Ratio=25%, Saving FLOPs. (%)=64.302025.11 | 77.14 | 69.21 | 9.95 | — | |
| UniDriveVLAParams=8B2026.04 | 76.3 | — | — | — | |
| MUSE-1BType=Unified, # Params=1B+0.6B2026.05 | 72.5 | — | — | — | |
| SparseVLMVision Token Retention Ratio=25%, Saving FLOPs. (%)=64.302025.11 | 72.12 | 54.73 | 24.43 | — | |
| UniLIP-1BType=Unified, # Params=1B+0.6B2026.05 | 70.7 | — | — | — | |
| GPT-5 Nano2026.04 | 65.7 | — | — | — |