SHARe-KAN: Post-Training Vector Quantization for Cache-Resident KAN Inference

ArXi:2512.15742v2 Announce Type: replace Pre-trained Vision Kolmogoro-Arnold Networks (KANs) a dense B-spline grid on every edge, inflating prediction-head parameter counts by than 140X relative to a comparable MLP and pushing inference into a memory-bound regime on edge accelerators. Standard magnitude pruning fails on these pre-trained models: zero-shot sparsity collapses accuracy, and restoring it requires an iterative fine-tuning loop that is impractical in deployment settings. We present SHARe-KAN, a post.