QuantKAN: A Unified Quantization Framework for Kolmogorov Arnold Networks

A new framework called QuantKAN provides the first unified approach for compressing Kolmogorov–Arnold Networks (KANs) to run efficiently on hardware, according to a paper posted on arXiv [1]. The work addresses a key barrier to deploying the spline-based models in low-precision environments. KANs differ from conventional neural networks by replacing linear weights with spline-based functions, a design that boosts expressivity but creates uneven parameter distributions that complicate quantization [1]. The QuantKAN framework tackles this with branch-aware quantizers that separately handle base and spline parameters, and it extends modern quantization-aware training (QAT) and post-training quantization (PTQ) techniques to spline-based layers across four KAN variants: EfficientKAN, FastKAN, PyKAN, and KAGN [1]. The paper, authored by researchers including Lizhong Chen, reports the first unified QAT and PTQ benchmarks for KANs on datasets ranging from MNIST to ImageNet [1]. Among QAT methods, Differentiable Soft Quantization (DSQ) proved most robust under aggressive low-bit constraints, while GPTQ emerged as the strongest PTQ method at moderate precision [1]. Sensitivity analyses uncovered architecture-specific failure modes: spline and basis parameters dominate in FastKAN, whereas base or scaling parameters are the critical points in EfficientKAN, GRAM, and PyKAN [1]. Hardware projections using Vivado HLS on a Xilinx UltraScale+ device suggest that a W4A4 configuration can deliver up to 3.32 times the throughput and 7.7 times lower estimated dynamic energy per inference compared to a baseline [1]. However, the analysis also identifies a residual “basis-evaluation tax” that persists even after quantization, prompting the authors to call for basis-aware microarchitecture design [1]. The QuantKAN codebase is publicly available on GitHub [1]. The paper has been revised twice since its initial submission in November 2025, with the latest version posted in June 2026 [1]. arXiv, the open-access repository hosting the work, was founded in 1991 and now receives roughly 24,000 submissions per month, serving as a primary distribution channel for preprints in computer science, physics, and mathematics [6].