Hugging Face Papers ↗ · 2026-11-05
Module-Aware Optimization for Graph Neural Networks
Media reports Hugging Face Papers Source language: Chinese
论文摘要(作者报告,未经本站复测):Different optimizers impose different geometries on parameter updates.
Full text
GraphMuon adapts Muon's matrix orthogonalization to the covariance of graph-propagated representations, reducing graph-induced update amplification and speeding validation-loss time-to-target on graph benchmarks.