Bridging Theory and Practice with KANX

Kolmogorov–Arnold Networks (KANs) are a new class of function-approximating neural models inspired by Kolmogorov’s superposition theorem [2]. In KANs, each network weight is a learnable univariate function (modeled as a spline) rather than a fixed linear coefficient [1]. Recent work shows that KANs can outperform multi-layer perceptrons (MLPs) on synthetic tasks, achieving much lower error with far fewer parameters [1][3]. However, a reproducible, production-ready KAN implementation and comprehensive experimental evaluation have been lacking. We present KANX, an open-source library (TensorFlow and PyTorch backends) that implements KANs in a clean, deployable format. KANX includes automated tests (92% coverage), CLI, FastAPI service, Docker/Kubernetes manifests, and continuous release (PyPI) [4]. We identify the research gap in existing KAN literature: limited real-world experimentation and no standard benchmarks or tools. We outline hypotheses to test (e.g., KANs’ scaling behavior vs. MLPs) and design experiments using synthetic regression tasks (e.g. y=sin⁡(πx1)+cos⁡(2πx2)) and standard datasets. Results from KANX’s benchmarks confirm that small KANs (e.g., [2, 32, 1]) achieve orders-of-magnitude lower MSE than similarly sized MLPs [3]. We provide reproducible code to replicate these findings and propose additional ablations (varying layer widths, spline order, etc.) to strengthen the analysis. Finally, we give publication recommendations (arXiv preprint, then ML venues) and required additions for each.  Our work formalizes KANX as both a research artifact and a state-of-the-art implementation of KANs, advancing the field by enabling fair comparisons and broader adoption