Visual Question Answering on GQA (Accuracy, RelAcc.)
64.2Accuracy (GQA)Vanilla
Evaluation results
| Method | Links | ||
|---|---|---|---|
| VanillaToken Retention Budget=2,880 Tokens (100%), Backbone=LLaVA-NeXT 7B2026.05 | 64.2 | 100 | |
| OccamTokenToken Retention Budget=320 Tokens (↓ 88.9%), Backbone=LLaVA-NeXT 7B2026.05 | 63 | 97.3 | |
| TwigVLMToken Retention Budget=320 Tokens (↓ 88.9%), Backbone=LLaVA-NeXT 7B, Params.=610M2026.05 | 62.2 | 96.3 | |
| OccamTokenToken Retention Budget=160 Tokens (↓ 94.4%), Backbone=LLaVA-NeXT 7B2026.05 | 61.3 | 95.9 | |
| DivPruneToken Retention Budget=320 Tokens (↓ 88.9%), Backbone=LLaVA-NeXT 7B2026.05 | 61.1 | 95 | |
| DARTToken Retention Budget=320 Tokens (↓ 88.9%), Backbone=LLaVA-NeXT 7B2026.05 | 59.5 | 93.8 | |
| OccamTokenToken Retention Budget=40 Tokens (↓ 98.6%), Backbone=LLaVA-NeXT 7B2026.05 | 59.5 | 93.8 | |
| VisionZipToken Retention Budget=320 Tokens (↓ 88.9%), Backbone=LLaVA-NeXT 7B2026.05 | 59.3 | 94 | |
| DivPruneToken Retention Budget=160 Tokens (↓ 94.4%), Backbone=LLaVA-NeXT 7B2026.05 | 59.3 | 91.7 | |
| DARTToken Retention Budget=160 Tokens (↓ 94.4%), Backbone=LLaVA-NeXT 7B2026.05 | 56.8 | 90.6 | |
| VisionZipToken Retention Budget=160 Tokens (↓ 94.4%), Backbone=LLaVA-NeXT 7B2026.05 | 55.5 | 89.4 | |
| DivPruneToken Retention Budget=40 Tokens (↓ 98.6%), Backbone=LLaVA-NeXT 7B2026.05 | 54 | 83.4 | |
| VisionZipToken Retention Budget=40 Tokens (↓ 98.6%), Backbone=LLaVA-NeXT 7B2026.05 | 42.2 | 63.8 |