{"id":"arxiv/2610.21077","url":"https://arxiv.org/abs/2610.21077","tags":["math.NA","cs.LG"],"score":0,"title":"Adaptive Randomized Neural Networks for PDEs: Structure-Aware Feature Redistribution and Residual-Driven Block Enrichment","domain":"math.NA","source":"arxiv","authors":["You Yang","Fei Wang"],"reviews":0,"subject":"math.NA","summary":"Randomized neural networks retain an output-linear least-squares solve, but globally sampled random features may fail to capture the spatial and scale structure of PDE solutions. We propose SFRBE--RaNN, a two-stage adaptive trial-space construction. SFR redistributes a fixed feature budget according to regional indicator masses and adapts regional weight scales through low-dimensional residual-based optimization. RBE constructs nonredundant enrichments by screening multiple activation families, removing represented response directions, and selecting the smallest admissible prefix. For…","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-10-15","reactions":{}}