Preventing Expert Collapse in MoE-dVLMs via Modality-Wise Norm Alignment

Benchmark Model Rank Results
chart-question-answering-on-chartqaTowerAlign–1:1 Accuracy: 59.0
multimodal-reasoning-on-mmmuTowerAlign–Accuracy: 40.3
vision-language-understanding-on-mmstarTowerAlign–Accuracy: 46.7
visual-question-answering-vqa-on-realworldqaTowerAlign–Accuracy: 58.9