AIBullisharXiv – CS AI · 5h ago7/10
🧠
E2Former-V2: On-the-Fly Equivariant Attention with Linear Activation Memory
Researchers introduce E2Former-V2, a more scalable architecture for Equivariant Graph Neural Networks that models 3D molecular systems. By combining algebraic sparsity with hardware-optimized execution, the model achieves 20× computational improvements while maintaining competitive accuracy on molecular datasets.