Reproduce “Pretraining Data Scale Reverses the Muon-AdamW Fine-Tuning Gap” 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
Catherine Li, Jingchu Gai, Andrej Risteski, Jacob Mitchell Springer, Aditi Raghunathan
The authors compare how AdamW and Muon pretraining affect language-model fine-tuning across five OLMo-style model sizes and four datasets. They measure learning–forgetting frontiers and track the margin-to-spread ratio of next-token logits. Muon leads at lower data budgets, but AdamW often overtakes it beyond task- and model-dependent crossovers, despite Muon’s lower pretraining loss; this does not establish a universal ordering.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.