Knowledge Preservation on VQAv2, GQA, VizWiz, SQA, TextVQA, POPE, MM-Bench, MM-Bench-CN
80.7VQAv2 AccuracyPre-trained Model
Evaluation results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Pre-trained ModelBackbone=Qwen-2.5VL-3B, Target Task=Flickr30k2026.06 | 80.7 | 58.9 | 65.9 | 82.7 | 75.6 | 87.6 | 77.7 | 76.6 | 75.7 | |
| CGMBackbone=Qwen-2.5VL-3B, Target Task=Flickr30k2026.06 | 79.6 | 56.1 | 63.5 | 82 | 73.2 | 86.9 | 78.1 | 76.8 | 74.5 | |
| CGM†Backbone=Qwen-2.5VL-3B, Target Task=Flickr30k2026.06 | 76.9 | 53 | 61 | 82.2 | 72.4 | 86.5 | 78.2 | 76.7 | 73.3 | |
| DAREBackbone=Qwen-2.5VL-3B, Target Task=Flickr30k2026.06 | 73.4 | 47.8 | 54.2 | 78.2 | 64 | 86.6 | 73.7 | 57 | 66.9 | |
| Fine-tuned ModelBackbone=Qwen-2.5VL-3B, Target Task=Flickr30k2026.06 | 72.9 | 46.8 | 57 | 79.2 | 66.2 | 83.9 | 75.6 | 63 | 68.1 | |
| WandaBackbone=Qwen-2.5VL-3B, Target Task=Flickr30k2026.06 | 72.8 | 46.8 | 57.2 | 79.3 | 66.1 | 83.9 | 75.4 | 63.1 | 68.1 | |
| MagnitudeBackbone=Qwen-2.5VL-3B, Target Task=Flickr30k2026.06 | 72 | 45.8 | 56.4 | 79.9 | 65.7 | 83.2 | 76.2 | 63.7 | 67.8 | |
| TailorBackbone=Qwen-2.5VL-3B, Target Task=Flickr30k2026.06 | 70.3 | 46.4 | 56.1 | 79.7 | 61.2 | 84.6 | 75.6 | 60.7 | 66.8 | |
| GraftingBackbone=Qwen-2.5VL-3B, Target Task=Flickr30k2026.06 | 68.8 | 44.3 | 59.3 | 80.5 | 62.4 | 76 | 76.7 | 70.1 | 67.2 |