In recent years, Muon has emerged as the dominant method for training large language models, and transformers more broadly. The essential difference, when compared to standard gradient descent methods, is to replace the usual update matrix M=UΣV^top with its polar factor UV^top. In this work, we consider a class of Muon-like updates, where we replace the update M with <strong>UΣ^p V^top</strong> for some parameter p. We call this a \"spectral-shaping\" operation, and develop a theory of how to pick p which depends on (a) local curvature of the loss function, (b) noise stemming from stochastic gradients and label noise, and (c) training stage. Our theory and experimentation reveal a previously overlooked behavior: positive p helps early by emphasizing high-curvature directions and accelerating signal contraction, while mildly negative p helps later by reallocating update strength toward low-curvature directions that still contain useful training signals. Building on the insight, we propose <strong>DynMuon</strong>, an efficient dynamic spectral shaping method that schedules p from positive to mildly negative over training. Extensive experiments across model sizes, architectures, and training settings show that DynMuon consistently achieves lower validation loss than Muon, while requiring 10.6-26.5% fewer steps to reach the same target loss.</p>\n","updatedAt":"2026-05-21T22:33:50.513Z","author":{"_id":"651f30ce9f372ea08ddc5b1c","avatarUrl":"/avatars/db6eb1e1f50477740a653529c4657039.svg","fullname":"Fangzhou Wu","name":"wark123","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.904931366443634},"editors":["wark123"],"editorAvatarUrls":["/avatars/db6eb1e1f50477740a653529c4657039.svg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2605.17109","authors":[{"_id":"6a0cbbd465eb30f20d9629e6","name":"Fangzhou Wu","hidden":false},{"_id":"6a0cbbd465eb30f20d9629e7","name":"Rikhav Shah","hidden":false},{"_id":"6a0cbbd465eb30f20d9629e8","name":"Sandeep Silwal","hidden":false},{"_id":"6a0cbbd465eb30f20d9629e9","name":"Qiuyi Zhang","hidden":false}],"publishedAt":"2026-05-16T00:00:00.000Z","submittedOnDailyAt":"2026-05-21T00:00:00.000Z","title":"DynMuon: A Dynamic Spectral Shaping View of Muon","submittedOnDailyBy":{"_id":"651f30ce9f372ea08ddc5b1c","avatarUrl":"/avatars/db6eb1e1f50477740a653529c4657039.svg","isPro":false,"fullname":"Fangzhou Wu","user":"wark123","type":"user","name":"wark123"},"summary":"In recent years, Muon has emerged as the dominant method for training large language models, and transformers more broadly. The essential difference, when compared to standard gradient descent methods, is to replace the usual update matrix M=UΣV^top with its polar factor UV^top. In this work, we consider a class of Muon-like updates, where we replace the update M with UΣ^p V^top for some parameter p. We call this a \"spectral-shaping\" operation, and develop a theory of how to pick p which depends on (a) local curvature of the loss function, (b) noise stemming from stochastic gradients and label noise, and (c) training stage. Our theory and experimentation reveal a previously overlooked behavior: positive p helps early by emphasizing high-curvature directions and accelerating signal contraction, while mildly negative p helps later by reallocating update strength toward low-curvature directions that still contain useful training signals. Building on the insight, we propose DynMuon, an efficient dynamic spectral shaping method that schedules p from positive to mildly negative over training. Extensive experiments across model sizes, architectures, and training settings show that DynMuon consistently achieves lower validation loss than Muon, while requiring 10.6-26.5% fewer steps to reach the same target loss.","upvotes":1,"discussionId":"6a0cbbd565eb30f20d9629ea","ai_summary":"Muon optimizer's spectral-shaping approach dynamically adjusts update parameters during training to improve convergence, achieving better validation loss with fewer training steps.","ai_keywords":["Muon","gradient descent","update matrix","polar factor","spectral-shaping","loss function","stochastic gradients","training stage","DynMuon","optimization"],"organization":{"_id":"6279cd50c20d41b28913755e","name":"Uwmadison","fullname":"University of Wisconsin-Madison","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/9Jg5WNG52u9RTFzOgQsII.png"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"651f30ce9f372ea08ddc5b1c","avatarUrl":"/avatars/db6eb1e1f50477740a653529c4657039.svg","isPro":false,"fullname":"Fangzhou Wu","user":"wark123","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"organization":{"_id":"6279cd50c20d41b28913755e","name":"Uwmadison","fullname":"University of Wisconsin-Madison","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/9Jg5WNG52u9RTFzOgQsII.png"}}">
DynMuon: A Dynamic Spectral Shaping View of Muon
Abstract
Muon optimizer's spectral-shaping approach dynamically adjusts update parameters during training to improve convergence, achieving better validation loss with fewer training steps.
AI-generated summary
In recent years, Muon has emerged as the dominant method for training large language models, and transformers more broadly. The essential difference, when compared to standard gradient descent methods, is to replace the usual update matrix M=UΣV^top with its polar factor UV^top. In this work, we consider a class of Muon-like updates, where we replace the update M with UΣ^p V^top for some parameter p. We call this a "spectral-shaping" operation, and develop a theory of how to pick p which depends on (a) local curvature of the loss function, (b) noise stemming from stochastic gradients and label noise, and (c) training stage. Our theory and experimentation reveal a previously overlooked behavior: positive p helps early by emphasizing high-curvature directions and accelerating signal contraction, while mildly negative p helps later by reallocating update strength toward low-curvature directions that still contain useful training signals. Building on the insight, we propose DynMuon, an efficient dynamic spectral shaping method that schedules p from positive to mildly negative over training. Extensive experiments across model sizes, architectures, and training settings show that DynMuon consistently achieves lower validation loss than Muon, while requiring 10.6-26.5% fewer steps to reach the same target loss.
Community
In recent years, Muon has emerged as the dominant method for training large language models, and transformers more broadly. The essential difference, when compared to standard gradient descent methods, is to replace the usual update matrix M=UΣV^top with its polar factor UV^top. In this work, we consider a class of Muon-like updates, where we replace the update M with UΣ^p V^top for some parameter p. We call this a "spectral-shaping" operation, and develop a theory of how to pick p which depends on (a) local curvature of the loss function, (b) noise stemming from stochastic gradients and label noise, and (c) training stage. Our theory and experimentation reveal a previously overlooked behavior: positive p helps early by emphasizing high-curvature directions and accelerating signal contraction, while mildly negative p helps later by reallocating update strength toward low-curvature directions that still contain useful training signals. Building on the insight, we propose DynMuon, an efficient dynamic spectral shaping method that schedules p from positive to mildly negative over training. Extensive experiments across model sizes, architectures, and training settings show that DynMuon consistently achieves lower validation loss than Muon, while requiring 10.6-26.5% fewer steps to reach the same target loss.
Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.
Tap or paste here to upload images
Cite arxiv.org/abs/2605.17109 in a model README.md to link it from this page.
Cite arxiv.org/abs/2605.17109 in a dataset README.md to link it from this page.
Cite arxiv.org/abs/2605.17109 in a Space README.md to link it from this page.
Discussion (0)
Sign in to join the discussion. Free account, 30 seconds — email code or GitHub.
Sign in →No comments yet. Sign in and be the first to say something.