Visual Question Answering on GQA (Accuracy and Relative Performance)
63.9AccuracyVanilla
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| VanillaToken Retention Ratio=100%, Backbone=LLaVA-NeXT-7B2026.02 | 63.9 | 100 | |
| SwiftVLMToken Retention Ratio=33.3%, Backbone=LLaVA-NeXT-7B2026.02 | 63.6 | 98 | |
| SwiftVLMToken Retention Ratio=22.2%, Backbone=LLaVA-NeXT-7B2026.02 | 63.5 | 97.1 | |
| LLaVA-1.5-VL-13BBackbone=LLaVA-1.5-VL-13B, Token=5762026.03 | 63.3 | 100 | |
| TPRLBackbone=LLaVA-1.5-VL-13B, Token=192*2026.03 | 62.8 | 99.2 | |
| FEATHERToken Retention Ratio=33.3%, Backbone=LLaVA-NeXT-7B2026.02 | 62.5 | 92.8 | |
| BaseBackbone=LLaVA-1.5-7B2026.03 | 62 | — | |
| FEATHERToken Retention Ratio=22.2%, Backbone=LLaVA-NeXT-7B2026.02 | 61.9 | 87.5 | |
| VanillaVisual Token Budget=5762026.02 | 61.4 | 100 | |
| TPRLBackbone=LLaVA-1.5-VL-13B, Token=128*2026.03 | 61.2 | 96.7 | |
| VisionZip+EEnergy Threshold (τ)=99.8%, Adaptive Budget (+E)=true, Backbone=LLaVA-1.5-7B, Mean Effective Token Count=3122026.03 | 60.9 | 98.2 | |
| VisionZipEnergy Threshold (τ)=99.8%, Adaptive Budget (+E)=false, Backbone=LLaVA-1.5-7B, Mean Effective Token Count=3122026.03 | 60.8 | 98.1 | |
| SwiftVLMVisual Token Budget=1922026.02 | 60.7 | 98.9 | |
| PDropEnergy Threshold (τ)=99.8%, Adaptive Budget (+E)=false, Backbone=LLaVA-1.5-7B, Mean Effective Token Count=3122026.03 | 60.5 | 97.6 | |
| PDrop+EEnergy Threshold (τ)=99.8%, Adaptive Budget (+E)=true, Backbone=LLaVA-1.5-7B, Mean Effective Token Count=3122026.03 | 60.5 | 97.6 | |
| FastVEnergy Threshold (τ)=99.8%, Adaptive Budget (+E)=false, Backbone=LLaVA-1.5-7B, Mean Effective Token Count=3122026.03 | 60.4 | 97.4 | |
| FastV+EEnergy Threshold (τ)=99.8%, Adaptive Budget (+E)=true, Backbone=LLaVA-1.5-7B, Mean Effective Token Count=3122026.03 | 60.3 | 97.3 | |
| VisionZip+EEnergy Threshold (τ)=99.0%, Adaptive Budget (+E)=true, Backbone=LLaVA-1.5-7B, Mean Effective Token Count=1772026.03 | 59.5 | 96 | |
| SwiftVLMVisual Token Budget=1282026.02 | 59.2 | 96.4 | |
| VisionZipEnergy Threshold (τ)=99.0%, Adaptive Budget (+E)=false, Backbone=LLaVA-1.5-7B, Mean Effective Token Count=1772026.03 | 59.1 | 95.3 | |
| FastVToken Retention Ratio=33.3%, Backbone=LLaVA-NeXT-7B2026.02 | 59 | 75.1 | |
| PDrop+EEnergy Threshold (τ)=99.0%, Adaptive Budget (+E)=true, Backbone=LLaVA-1.5-7B, Mean Effective Token Count=1772026.03 | 58.3 | 94 | |
| PDropEnergy Threshold (τ)=99.0%, Adaptive Budget (+E)=false, Backbone=LLaVA-1.5-7B, Mean Effective Token Count=1772026.03 | 58.1 | 93.7 | |
| FastV+EEnergy Threshold (τ)=99.0%, Adaptive Budget (+E)=true, Backbone=LLaVA-1.5-7B, Mean Effective Token Count=1772026.03 | 57.5 | 92.7 | |
| FastVEnergy Threshold (τ)=99.0%, Adaptive Budget (+E)=false, Backbone=LLaVA-1.5-7B, Mean Effective Token Count=1772026.03 | 57.4 | 92.6 | |
| FastVVisual Token Budget=1922026.02 | 57.2 | 93.2 | |
| FastVToken Retention Ratio=22.2%, Backbone=LLaVA-NeXT-7B2026.02 | 56.9 | 66.9 | |
| FastVBackbone=LLaVA-1.5-VL-13B, Token=1922026.03 | 54.2 | 85.7 | |
| FastVBackbone=LLaVA-1.5-VL-13B, Token=1282026.03 | 48.5 | 76.6 |