Image Question Answering on GQA, SQAImg, and TextVQA
59.1Accuracy (GQA)EchoPrune
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| EchoPruneTokens Kept=128, Token Retention Ratio=22.2%, Backbone=LLaVA1.5-7B2026.05 | 59.1 | 69.7 | 57.1 | 62 | |
| ApETTokens Kept=128, Token Retention Ratio=22.2%, Backbone=LLaVA1.5-7B2026.05 | 58.9 | 68.7 | 53.9 | 60.5 | |
| MI-PrunerTokens Kept=128, Token Retention Ratio=22.2%, Backbone=LLaVA1.5-7B2026.05 | 58.5 | 69.5 | 56.1 | 61.4 | |
| VisionZipTokens Kept=128, Token Retention Ratio=22.2%, Backbone=LLaVA1.5-7B2026.05 | 57.6 | 68.9 | 56.8 | 61.1 | |
| EchoPruneTokens Kept=64, Token Retention Ratio=11.1%, Backbone=LLaVA1.5-7B2026.05 | 57.6 | 68.5 | 55.6 | 60.6 | |
| MI-PrunerTokens Kept=64, Token Retention Ratio=11.1%, Backbone=LLaVA1.5-7B2026.05 | 56.9 | 69.8 | 54.9 | 60.5 | |
| ApETTokens Kept=64, Token Retention Ratio=11.1%, Backbone=LLaVA1.5-7B2026.05 | 56.9 | 68.9 | 53 | 59.6 | |
| VisionZipTokens Kept=64, Token Retention Ratio=11.1%, Backbone=LLaVA1.5-7B2026.05 | 55.1 | 69 | 55.5 | 59.9 |