Reproduce “Muon Operates Beyond the Edge of Stability” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Niko Covic, Baraq Lipshitz, Alexander Padula, Stefan Visnjic

The authors measure how Muon’s matrix-based updates relate to sharpness, generalization, and training stability. In fixed-learning-rate CIFAR-10 runs, adaptive sharpness correlates with the generalization gap within optimizer groups, including Decoupled Muon (Pearson r = 0.51), but not under linear decay. In a full-batch MSE experiment, Muon exceeds the expected sharpness threshold without loss divergence; this finding is limited to that setup.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.