tl;dr: Researchers designed FPGA architectures for Kolmogorov-Arnold Networks (KANs) that exploit their summed univariate activations, mapping each activation to a lookup table for sub-microsecond inference with a 2700x speedup over prior KAN-FPGA implementations. They further leverage B-spline locality (only S+1 basis functions active per input) and boundedness (basis functions sum to 1) to enable sparse, stable fixed-point gradient updates, supporting real-time on-FPGA online learning at 50,000+ parameters with sub-microsecond forward/backward passes—previously considered impractical. Applications include quantum control and nuclear fusion, where models must adapt within microseconds.