Geometry-constrained KANs are significantly more robust to measurement noise than unregularized splines, degrading only 3.7x compared to 21.6x as noise increases from 0 to 1

as $σ$ grows from $0$ to $1$, $\ell^p$-KAN degrades only $3.7\times$ -- below even a cross-validated spline ($\approx 11\times$) -- while an unregularised spline degrades $21.6\times$

http://arxiv.org/abs/2608.25807v1