Text-intensive Visual Question Answering on VQAText
61.3ScoreVanilla
Evaluation Results
| Method | Links | |
|---|---|---|
| VanillaBackbone=LLaVA-NeXT-7B, Retained Tokens=2880, Token Retention Ratio=100%2026.06 | 61.3 | |
| VisionZipBackbone=LLaVA-NeXT-7B, Retained Tokens=640, Token Retention Ratio=↓ 77.8%2026.06 | 60.2 | |
| EvoCutBackbone=LLaVA-NeXT-7B, Retained Tokens=640, Token Retention Ratio=↓ 77.8%2026.06 | 59.7 | |
| VisionZipBackbone=LLaVA-NeXT-7B, Retained Tokens=320, Token Retention Ratio=↓ 88.9%2026.06 | 58.9 | |
| LBR (ours)Backbone=LLaVA-NEXT, Parameter Scale=3B, LBR Mitigation=true2026.05 | 57.9 | |
| ApETBackbone=LLaVA-NeXT-7B, Retained Tokens=640, Token Retention Ratio=↓ 77.8%2026.06 | 57.9 | |
| EvoCutBackbone=LLaVA-NeXT-7B, Retained Tokens=320, Token Retention Ratio=↓ 88.9%2026.06 | 57.9 | |
| PDropBackbone=LLaVA-NeXT-7B, Retained Tokens=640, Token Retention Ratio=↓ 77.8%2026.06 | 57.4 | |
| EvoCutBackbone=LLaVA-NeXT-7B, Retained Tokens=160, Token Retention Ratio=↓ 94.4%2026.06 | 57.3 | |
| VisionZipBackbone=LLaVA-NeXT-7B, Retained Tokens=160, Token Retention Ratio=↓ 94.4%2026.06 | 56.2 | |
| LLaVA-NEXT-3BBackbone=LLaVA-NEXT, Parameter Scale=3B, LBR Mitigation=false2026.05 | 56.1 | |
| PDropBackbone=LLaVA-NeXT-7B, Retained Tokens=320, Token Retention Ratio=↓ 88.9%2026.06 | 54.4 | |
| ApETBackbone=LLaVA-NeXT-7B, Retained Tokens=320, Token Retention Ratio=↓ 88.9%2026.06 | 54.4 | |
| ApETBackbone=LLaVA-NeXT-7B, Retained Tokens=160, Token Retention Ratio=↓ 94.4%2026.06 | 53.8 | |
| PDropBackbone=LLaVA-NeXT-7B, Retained Tokens=160, Token Retention Ratio=↓ 94.4%2026.06 | 52.7 | |
| LBR (ours)Backbone=LLaVA-1.5, Parameter Scale=13B, LBR Mitigation=true2026.05 | 48.1 | |
| LLaVA-1.5-13BBackbone=LLaVA-1.5, Parameter Scale=13B, LBR Mitigation=false2026.05 | 48 | |
| LBR (ours)Backbone=LLaVA-1.5, Parameter Scale=7B, LBR Mitigation=true2026.05 | 46 | |
| LLaVA-1.5-7BBackbone=LLaVA-1.5, Parameter Scale=7B, LBR Mitigation=false2026.05 | 45.8 |