Multimodal Visual Question Answering on VLMEvalKit Image Benchmarks
46.8HallBench AccuracyQwen2.5-VL-7B
Evaluation results
| Method | Links | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Qwen2.5-VL-7BTokens Retained=1296, Pruning Ratio=0%, Backbone=Qwen2.5-VL-7B, Input Resolution=1008x10082026.04 | 46.8 | 2,322 | 85.1 | 86.3 | 80.3 | 66.3 | 58 | 70.7 | 72.2 | 85.3 | 100 | |
| HAWKTokens Retained=512, Pruning Ratio=60.5%, Backbone=Qwen2.5-VL-7B, Input Resolution=1008x10082026.04 | 44.1 | 2,342 | 85.3 | 84.1 | 79.6 | 65.4 | 57.3 | 71 | 72.6 | 85 | 99 | |
| DivPruneTokens Retained=512, Pruning Ratio=60.5%, Pruning Strategy=Similarity-based, Backbone=Qwen2.5-VL-7B, Input Resolution=1008x10082026.04 | 43.5 | 2,258 | 81.8 | 80.3 | 78.7 | 63.8 | 55.1 | 69 | 71 | 84.4 | 96.4 | |
| CDPrunerTokens Retained=512, Pruning Ratio=60.5%, Pruning Strategy=Similarity & Attention-based, Backbone=Qwen2.5-VL-7B, Input Resolution=1008x10082026.04 | 42.5 | 2,327 | 84.2 | 82.8 | 78.9 | 64.8 | 55.7 | 67.1 | 71.3 | 82.6 | 96.8 | |
| FastVTokens Retained=512, Pruning Ratio=60.5%, Pruning Strategy=Attention-based, Backbone=Qwen2.5-VL-7B, Input Resolution=1008x10082026.04 | 42.4 | 2,318 | 84.1 | 82.7 | 78.8 | 64.7 | 55.8 | 67.8 | 71.5 | 84 | 97 | |
| HAWKTokens Retained=256, Pruning Ratio=80.2%, Backbone=Qwen2.5-VL-7B, Input Resolution=1008x10082026.04 | 42.4 | 2,245 | 83.7 | 77.2 | 77.7 | 63.7 | 55.5 | 70.3 | 71.9 | 82.9 | 96 | |
| HAWKTokens Retained=128, Pruning Ratio=90.1%, Backbone=Qwen2.5-VL-7B, Input Resolution=1008x10082026.04 | 40.5 | 2,137 | 79.2 | 64.4 | 75 | 59.4 | 49.4 | 67.8 | 70.2 | 77.8 | 90 | |
| CDPrunerTokens Retained=256, Pruning Ratio=80.2%, Pruning Strategy=Similarity & Attention-based, Backbone=Qwen2.5-VL-7B, Input Resolution=1008x10082026.04 | 40.1 | 2,245 | 82.4 | 73 | 77.5 | 62.2 | 54.3 | 61 | 70.7 | 79.4 | 92.5 | |
| DivPruneTokens Retained=256, Pruning Ratio=80.2%, Pruning Strategy=Similarity-based, Backbone=Qwen2.5-VL-7B, Input Resolution=1008x10082026.04 | 39.8 | 2,162 | 75.8 | 69 | 76.4 | 60.5 | 53.5 | 65.4 | 70.1 | 83.5 | 91.3 | |
| FastVTokens Retained=256, Pruning Ratio=80.2%, Pruning Strategy=Attention-based, Backbone=Qwen2.5-VL-7B, Input Resolution=1008x10082026.04 | 39 | 2,233 | 81.5 | 71.4 | 76.2 | 62.5 | 54.1 | 65.6 | 70.9 | 81.2 | 92.6 | |
| CDPrunerTokens Retained=128, Pruning Ratio=90.1%, Pruning Strategy=Similarity & Attention-based, Backbone=Qwen2.5-VL-7B, Input Resolution=1008x10082026.04 | 37.2 | 2,127 | 77.8 | 59.2 | 74 | 59.1 | 49.4 | 54.3 | 69.3 | 75.6 | 85.9 | |
| FastVTokens Retained=128, Pruning Ratio=90.1%, Pruning Strategy=Attention-based, Backbone=Qwen2.5-VL-7B, Input Resolution=1008x10082026.04 | 33.8 | 2,028 | 73.8 | 54.3 | 71.4 | 57.8 | 46.7 | 62.7 | 69.4 | 75.8 | 83.9 | |
| DivPruneTokens Retained=128, Pruning Ratio=90.1%, Pruning Strategy=Similarity-based, Backbone=Qwen2.5-VL-7B, Input Resolution=1008x10082026.04 | 33.4 | 1,979 | 66.6 | 52.5 | 72.1 | 57.9 | 48.4 | 57.8 | 68.6 | 81.8 | 82.8 |