Multimodal Understanding on GQA, MMB, MMB-CN, MME, POPE, SQA, and TextVQA
63.1GQA AccuracyVanilla
Evaluation Results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| VanillaBackbone=InternVL2.5-8B, Token Retention Ratio=100%, Visual Token Count=1792 Tokens, Evaluation Protocol=Upper Bound2026.06 | 63.1 | 84.7 | 82.6 | 2,332 | 90.5 | 98 | 76.8 | 100 | |
| FastVBackbone=InternVL2.5-8B, Token Retention Ratio=33.3%, Evaluation Protocol=Retain Tokens2026.06 | 61.1 | 83.3 | 81.5 | 2,251 | 90.3 | 97.3 | 75.1 | 98.2 | |
| PriorTRBackbone=InternVL2.5-8B, Token Retention Ratio=33.3%, Evaluation Protocol=Retain Tokens2026.06 | 61 | 83.8 | 83.1 | 2,275 | 90.5 | 97.3 | 73.5 | 98.4 | |
| PriorTRBackbone=InternVL2.5-8B, Token Retention Ratio=22.2%, Evaluation Protocol=Retain Tokens2026.06 | 60.2 | 82.7 | 81.4 | 2,263 | 90.8 | 95.7 | 71.9 | 97.2 | |
| FastVBackbone=InternVL2.5-8B, Token Retention Ratio=22.2%, Evaluation Protocol=Retain Tokens2026.06 | 59.6 | 81 | 80.3 | 2,176 | 89.6 | 94.9 | 73.3 | 96 | |
| PriorTRBackbone=InternVL2.5-8B, Token Retention Ratio=11.1%, Evaluation Protocol=Retain Tokens2026.06 | 56.8 | 80.8 | 78.7 | 2,150 | 89 | 91.2 | 67.5 | 93.2 | |
| FastVBackbone=InternVL2.5-8B, Token Retention Ratio=11.1%, Evaluation Protocol=Retain Tokens2026.06 | 53.1 | 74 | 72.8 | 1,970 | 84.7 | 88 | 65.3 | 87.5 |