CASA: Cross-Attention via Self-Attention for Efficient Vision-Language Fusion

Benchmark Model Rank Results
video-question-answering-on-mlvuCAt+v (Qwen2.5-VL, 3B)#42Accuracy: 67.9
optical-character-recognition-ocr-on-ocrbenchCAQ2.5-VL#55Score: 783.0
chart-question-answering-on-chartqaCAt+v (Qwen2.5-VL, 3B)#641:1 Accuracy: 81.3
video-question-answering-on-mvbenchCAt+v (Qwen2.5-VL, 3B)#65Avg.: 60.6
visual-question-answering-vqa-on-realworldqaCAt+v (Qwen2.5-VL, 3B)#74Accuracy: 63.8