Functional Attention: From Pairwise Affinities to Functional Correspondences

Researchers introduced Functional Attention, a new attention mechanism for operator learning that models global dependencies with structured linear operators instead of token-wise softmax.

The method reframes attention as correspondence between adaptive bases, aiming to better represent continuous function spaces and resolution-invariant behavior.

The arXiv paper reports competitive results against strong baselines across operator learning benchmarks, including PDE solving, 3D segmentation, and regression.

Its compact, discretization-robust design could help scientific computing and vision models capture global structure more effectively.