Visual Question Answering Suite (VQAv2, GQA, VizWiz, SQA, TextVQA, POPE, InfoVQA, OKVQA)
80.9VQAv2 AccuracyCGM
Evaluation results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| CGMBackbone=Qwen-2.5VL-3B2026.06 | 80.9 | 58.9 | 66 | 82.7 | 75.8 | 87.6 | 60 | 57.1 | 71.1 | |
| Pre-trained ModelBackbone=Qwen-2.5VL-3B2026.06 | 80.7 | 58.9 | 65.9 | 82.7 | 75.6 | 87.6 | 61.3 | 56.1 | 71.1 | |
| CGM†Backbone=Qwen-2.5VL-3B2026.06 | 80.4 | 57.8 | 65 | 82.7 | 75.5 | 86.6 | 60.1 | 56.6 | 70.6 | |
| GraftingBackbone=Qwen-2.5VL-3B2026.06 | 72.6 | 48.4 | 53.8 | 82.7 | 72.3 | 83.9 | 59 | 48.3 | 65.1 | |
| TailorBackbone=Qwen-2.5VL-3B2026.06 | 69.7 | 45.5 | 54.4 | 82.3 | 70.4 | 82.3 | 58.2 | 46.7 | 63.7 | |
| MagnitudeBackbone=Qwen-2.5VL-3B2026.06 | 68.5 | 44.3 | 53.5 | 82.4 | 67.6 | 81.4 | 55.7 | 45.3 | 62.4 | |
| WandaBackbone=Qwen-2.5VL-3B2026.06 | 67.1 | 43.1 | 52.9 | 82.2 | 66.4 | 81 | 55.1 | 44.3 | 61.5 | |
| Fine-tuned ModelBackbone=Qwen-2.5VL-3B2026.06 | 66.8 | 42.9 | 52.7 | 81.8 | 65.9 | 80.8 | 54.8 | 44.1 | 61.2 | |
| DAREBackbone=Qwen-2.5VL-3B2026.06 | 66.3 | 44.1 | 50.9 | 78.5 | 63.3 | 82.9 | 53.1 | 42.7 | 60.2 |