Visual Question Answering on TextVQA (Accuracy, Rel. Acc)
64.7AccuracyDART
Evaluation results
| Method | Links | ||
|---|---|---|---|
| DARTRetained Tokens=640, Token Reduction Ratio=77.8%, Backbone=LLaVA-NeXT-13B2026.06 | 64.7 | 100.7 | |
| LLaVA-NeXT-13BRetained Tokens=2880, Token Reduction Ratio=100%, Backbone=LLaVA-NeXT-13B2026.06 | 63.2 | 100 | |
| SPARERetained Tokens=640, Token Reduction Ratio=77.8%, Backbone=LLaVA-NeXT-13B2026.06 | 61.1 | 99.1 | |
| CDPrunerRetained Tokens=640, Token Reduction Ratio=77.8%, Backbone=LLaVA-NeXT-13B2026.06 | 61 | 98.3 | |
| SPARERetained Tokens=320, Token Reduction Ratio=88.9%, Backbone=LLaVA-NeXT-13B2026.06 | 59.3 | 97.2 | |
| DARTRetained Tokens=320, Token Reduction Ratio=88.9%, Backbone=LLaVA-NeXT-13B2026.06 | 58.8 | 96.3 | |
| CDPrunerRetained Tokens=320, Token Reduction Ratio=88.9%, Backbone=LLaVA-NeXT-13B2026.06 | 58.8 | 97.1 | |
| SPARERetained Tokens=160, Token Reduction Ratio=94.4%, Backbone=LLaVA-NeXT-13B2026.06 | 57.9 | 95.6 | |
| CDPrunerRetained Tokens=160, Token Reduction Ratio=94.4%, Backbone=LLaVA-NeXT-13B2026.06 | 56.7 | 95.2 | |
| DARTRetained Tokens=160, Token Reduction Ratio=94.4%, Backbone=LLaVA-NeXT-13B2026.06 | 49.6 | 89.4 |