TL;DR of the paper, generated automatically.

New graph neural network optimizer, GraphMuon, claims faster training than AdamW.

• Idea: adapts Muon's matrix updates to graph-propagated features.
• Speed: 1.32x-2.42x faster validation-loss time-to-target than AdamW on three OGB benchmarks.
• Results: faster in 13 of 15 configurations, with competitive or superior predictive performance.