Video Understanding on VideoMME (Overall, Without Subtitles, and Relative Metrics)
58.4VideoMME Overall ScoreVanilla
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| VanillaToken Budget=All Tokens (100%), Base Model=LLaVA-OneVision-7B2026.05 | 58.4 | 61.2 | 100 | |
| SPprunerToken Budget=Retain 35% Tokens (↓ 65%), Base Model=LLaVA-OneVision-7B2026.05 | 57.9 | 60.2 | 98.9 | |
| SPprunerToken Budget=Retain 25% Tokens (↓ 75%), Base Model=LLaVA-OneVision-7B2026.05 | 57.2 | 60.4 | 98.7 | |
| FastVToken Budget=Retain 35% Tokens (↓ 65%), Base Model=LLaVA-OneVision-7B2026.05 | 57.1 | 59.8 | 97.6 | |
| SPprunerToken Budget=Retain 15% Tokens (↓ 85%), Base Model=LLaVA-OneVision-7B2026.05 | 56.2 | 58.7 | 96.7 | |
| FastVToken Budget=Retain 25% Tokens (↓ 75%), Base Model=LLaVA-OneVision-7B2026.05 | 56 | 58.7 | 96.3 | |
| FastVToken Budget=Retain 15% Tokens (↓ 85%), Base Model=LLaVA-OneVision-7B2026.05 | 53.6 | 57.6 | 93.7 |