A new arXiv paper takes a closer look at the Muon optimizer, whose advertised data-efficiency gains are often measured without considering real-world constraints. The authors propose a paired benchmark that explicitly controls for data, time, and tuning budgets, then compare a reference single-device Muon implementation against AdamW on a single A100 GPU.
According to the abstract, the evaluation spans model sizes starting at 17.7M parameters, though the full list is cut off in the summary. The key innovation is the "paired" setup: rather than letting each optimizer run with its own informal resource envelope, the study forces both to operate under the same declared limits.
The paper does not claim Muon always wins; instead, it asks under which budget conditions Muon's data efficiency actually translates into a practical advantage. Since the source is a single abstract, the article cannot yet report final numbers or conclusions beyond the methodology and motivation. Those details presumably appear in the full text on arXiv.