Spectral analysis reveals why Muon outpaces Adam in LLM training
Researchers have decoded why Muon, an orthogonal optimizer, trains large language models faster than Adam by analyzing loss landscapes across real training runs. Using spectral decomposition of momentum buffers on held-out data, they discovered that gradient directions exhibit a stable, anisotropic profile: a volatile high-frequency component constrains step sizes while a tolerant bulk permits aggressive updates. This unified framework spans model scales and optimizer families, offering a mechanistic foundation for optimizer design. The finding matters because it bridges the empirical success of second-order methods with interpretable theory, potentially enabling better pretraining efficiency across the industry.
Modelwire context
ExplainerThe paper doesn't just show Muon wins on speed; it isolates the specific loss landscape property that explains why: a two-tier gradient structure where high-frequency noise constrains step sizes while the bulk permits aggression. This mechanistic layer is absent from most optimizer comparisons, which typically report wall-clock time without explaining the underlying geometry.
This connects directly to the LoRA rank-error bounds paper from the same week. Both papers move beyond empirical trial-and-error toward principled, geometry-aware guidance: one derives error guarantees tied to task structure, the other decodes optimizer behavior through spectral properties of real training runs. Together they signal a shift in the field from 'what works' to 'why it works and when to apply it.' The Muon analysis also complements the test-time scaling work on token management; both expose inefficiencies in how we allocate compute (optimizer steps vs. inference tokens) by examining the actual structure of the problem rather than brute-force scaling.
If teams implementing Muon report consistent 15-20% pretraining speedups on models under 7B parameters within the next two quarters, the spectral framework holds across diverse architectures. If speedups flatten or reverse above 70B, that signals the anisotropic profile breaks down at scale and the theory needs refinement. Watch whether any major lab (Anthropic, DeepSeek, xAI) publicly adopts Muon in production pretraining and reports the measured gap versus their Adam baseline.
Coverage we drew on
This analysis is generated by Modelwire’s editorial layer from our archive and the summary above. It is not a substitute for the original reporting. How we write it.
MentionsMuon · Adam · Transformer
Modelwire Editorial
This synthesis and analysis was prepared by the Modelwire editorial team. We use advanced language models to read, ground, and connect the day’s most significant AI developments, providing original strategic context that helps practitioners and leaders stay ahead of the frontier.
Modelwire summarizes, we don’t republish. arXiv cs.LG originally reported this story as “Spectral Allocation: Why Muon Outperforms Adam, and How to Improve Muon”. 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.