Reproduce “Tune the Baseline First: Most of Muon's Scaled Weight Decay Gain Is a Tuning Gap, but Its Small Real Advantage Grows With Model Size” 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
Esther Xin
The authors compare Muon-SW and constant weight decay in language-model training, tuning each method’s learning-rate/weight-decay product separately. At 1.89 million active parameters, tuning reduces the held-out bits-per-byte gap from 0.03499 to 0.00261; all four paired seeds favor Muon-SW, but the residual is below the practical-significance threshold. Across three tested sizes, the surviving share rises from about 10% to 21%, though larger-size estimates use single seeds.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.