A new twist on Kolmogorov-Arnold Networks claims wins in two separate lanes, not one sweeping victory.
Researchers built SechKAN, a KAN variant that replaces the usual spline-based activation functions with a hyperbolic secant curve, a smooth, bell-shaped function chosen for well-behaved gradients. To keep the parameter count down, the team added a 1D linear projection, bringing SechKAN's model size in line with standard multilayer perceptrons (MLPs). They tested it on function fitting, PDE surrogate modeling, and image classification (MNIST, FashionMNIST, CIFAR-10, CIFAR-100). On PDE surrogate tasks, SechKAN beat MLPs and matched or beat other KAN variants. On image classification, it topped every other KAN variant tested while staying competitive with MLPs. Code is public on GitHub.
KANs have been pitched as an alternative to MLPs since 2024, moving learnable functions onto the network's edges instead of using fixed activations. The pitch has always come with an asterisk: KANs tend to cost more compute, which has kept them out of production systems. SechKAN does not erase that asterisk, but it narrows it in specific benchmarks rather than claiming a universal win, which is a more honest framing than most KAN papers manage.
The authors admit SechKAN is still pricier to run than MLPs and some KAN variants. A smoother curve is nice, but it does not fix the compute bill.