An arXiv preprint tests geometry-constrained KANs, finding strong clean-data results but different performance under noise, small data and extrapolation.