Image-text understanding on ScienceQA
88.7AccuracyPromPrune
Evaluation results
| Method | Links | ||
|---|---|---|---|
| PromPruneToken Budget=512, Reduction Percentage=60.5%2026.03 | 88.7 | 94.8 | |
| Qwen2.5-VL 7BToken Budget=All 1296 Tokens, Reduction Percentage=0%2026.03 | 88.6 | 100 | |
| VisPrunerToken Budget=512, Reduction Percentage=60.5%2026.03 | 88.5 | 94 | |
| DivPruneToken Budget=512, Reduction Percentage=60.5%2026.03 | 87.9 | 94.2 | |
| PromPruneToken Budget=256, Reduction Percentage=80.2%2026.03 | 86.4 | 92.6 | |
| VisPrunerToken Budget=256, Reduction Percentage=80.2%2026.03 | 86.1 | 91.7 | |
| DivPruneToken Budget=256, Reduction Percentage=80.2%2026.03 | 86.1 | 91.1 | |
| PromPruneToken Budget=128, Reduction Percentage=90.1%2026.03 | 84.5 | 87.9 | |
| DivPruneToken Budget=128, Reduction Percentage=90.1%2026.03 | 84.4 | 86.5 | |
| VisPrunerToken Budget=128, Reduction Percentage=90.1%2026.03 | 84.2 | 86.2 | |
| LLaVA-NeXT 7BToken Budget=2880, Base Model=LLaVA-NeXT 7B2026.03 | 68.6 | 100 | |
| VisPrunerToken Budget=160, Base Model=LLaVA-NeXT 7B2026.03 | 68.3 | 90.6 | |
| PromPruneToken Budget=640, Base Model=LLaVA-NeXT 7B2026.03 | 68.2 | 99.6 | |
| VisPrunerToken Budget=640, Base Model=LLaVA-NeXT 7B2026.03 | 68.1 | 99.2 | |
| DivPruneToken Budget=640, Base Model=LLaVA-NeXT 7B2026.03 | 68 | 99.6 | |
| PromPruneToken Budget=320, Base Model=LLaVA-NeXT 7B2026.03 | 67.8 | 97.9 | |
| VisPrunerToken Budget=320, Base Model=LLaVA-NeXT 7B2026.03 | 67.7 | 95.9 | |
| DivPruneToken Budget=320, Base Model=LLaVA-NeXT 7B2026.03 | 67.4 | 97.7 | |
| PromPruneToken Budget=160, Base Model=LLaVA-NeXT 7B2026.03 | 67.4 | 95.2 | |
| DivPruneToken Budget=160, Base Model=LLaVA-NeXT 7B2026.03 | 67.3 | 95 |