Muon's pretraining edge erodes at scale during fine-tuning
Source published ·Modelwire updated
Original coverage: arXiv cs.LG ↗·How Modelwire adds context
The development
Muon, an emerging optimizer for LLM pretraining, shows speed advantages over AdamW during training but loses ground after fine-tuning when pretraining data scales sufficiently. The reversal intensifies as model size grows, suggesting the relationship between optimizer choice and downstream task performance is more nuanced than previously understood. Researchers attribute this to how each optimizer affects logit margin dynamics during adaptation, introducing a margin-to-spread ratio metric to explain the phenomenon. This finding matters for practitioners choosing optimizers across the full training pipeline and hints that pretraining efficiency gains may not translate uniformly to production deployment.
Modelwire’s AI-generated summary of coverage from arXiv cs.LG.
Modelwire analysis
ExplainerOur AI-generated reading of the wider context and the next developments to watch.
The paper's real contribution isn't that Muon loses ground after fine-tuning, but that this reversal is data-scale dependent. At small pretraining scales Muon stays competitive; only at sufficient data volume does AdamW pull ahead. This inverts the usual framing: optimizer choice isn't universal, it's contingent on your pretraining budget.
This connects directly to the recent work on post-training trade-offs. The LLMON paper (late October) showed how architectural choices during training reshape downstream behavior; this Muon result extends that principle to optimizer selection. More relevant: the pass@k analysis from late October proved that single-attempt metrics hide real capability shifts. Here, the margin-to-spread ratio metric does similar work, revealing that what looks like fine-tuning degradation may reflect how each optimizer distributes logit probability rather than actual reasoning loss. The finding also echoes the pruning paper's core insight: apparent performance drops often mask different failure modes, not fundamental capability loss.
If practitioners report that Muon remains preferable for production fine-tuning on models pretrained below 2T tokens, the scale threshold claimed here is wrong. Conversely, if the next major open model trained with Muon (likely from a lab that adopted it post-2024) shows the predicted AdamW crossover during instruction tuning, that validates the scale dependency and signals optimizer choice should be explicitly tied to pretraining data budgets in deployment checklists.
This interpretation is generated from the summary above and the archive coverage cited below. Our methodology · Report an error
Coverage behind this analysis
These archive entries ground the connection in our analysis. They are ordered by source publication date, with links to our coverage and the original sources.
·arXiv cs.LG
RLVR training reveals unavoidable tradeoff between single and multi-attempt success
A new theoretical result quantifies a fundamental tradeoff in reinforcement learning for language models: optimizing for single-attempt success (pass@1) necessarily degrades performance when models get multiple chances to solve a problem. Researchers prove this isn't a training artifact but an inherent cost, modeling RLVR as an exponential tilt over shared strategies and using Stiemke duality…
MentionsMuon · AdamW
How this coverage is produced
Modelwire uses AI to generate summaries and context from source headlines, snippets, and selected archive coverage. Automated checks do not verify every claim, and items are not routinely reviewed by a person before publication. Zacaria Solis operates the site. Read the linked source for the full evidence and report errors through our corrections process.
Modelwire summarizes, we don’t republish. arXiv cs.LG originally reported this story as “Pretraining Data Scale Reverses the Muon-AdamW Fine-Tuning Gap”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.