Kun Ding

Pretrained visual encoders are a temporary crutch: encoder-free multimodal architectures close the performance gap with encoder-based systems at roughly $10^{22}$ FLOPs, well within current frontier pretraining budgets.

Achieving 94.52% of state-of-the-art performance with just 25% of the training data reveals a groundbreaking efficiency in visual instruction tuning.

DiffPDE achieves faster and more accurate PDE code repair by focusing only on localized errors, challenging the inefficiencies of existing autoregressive methods.

Depth integration in audio-visual segmentation leads to over 10% performance gains, revealing a critical yet overlooked modality in multimodal perception.