Multimodal VQA on LLaVA 1.5 Subset (GQA, MMB, POPE, MME) (test)
100AccuracyVanilla (LLaVA)
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Vanilla (LLaVA)Tokens=576, Backbone=LLaVA-v1.5-7B, Hardware=NVIDIA RTX A60002026.06 | 100 | 8.89 | 149.5 | 1 | |
| StepPruneTokens=64, K=64, Backbone=LLaVA-v1.5-7B, Hardware=NVIDIA RTX A60002026.06 | 94.6 | 2.28 | 79.5 | 1.88 | |
| PDropTokens=64, K=64, Backbone=LLaVA-v1.5-7B, Hardware=NVIDIA RTX A60002026.06 | 82.7 | 4.89 | 105.9 | 1.41 | |
| FastVTokens=64, K=64, Backbone=LLaVA-v1.5-7B, Hardware=NVIDIA RTX A60002026.06 | 76.8 | 2.39 | 102.6 | 1.46 |