Overview

Muon is a matrix-aware optimizer that accumulates momentum for weight matrices and approximately orthogonalizes each update, often with Newton–Schulz iterations, rather than scaling each parameter independently as Adam does. Research explores variants with curvature or variance preconditioning, different spectral weighting and normalization, and cheaper or distributed orthogonalization. Studies test Muon in language-model training, vision, recommendation, diffusion, and scientific machine learning, while theory examines convergence, noise, implicit bias, and generalization. A central open question is when its spectral updates improve learning enough to justify their cost: results depend on architecture, batch size, and training phase, and practical approximations can change both behavior and guarantees.

Latest papers

159 papers26 papers in the last four weeks, up 13% on the four weeks before. 0.2% of all new papers.

Oct 5, 2026

Does Muon Need Fine-Grained Spectral Shaping?

Muon combines current and past gradients into matrix momentum. For M=UΣV⊤M=UΣV^\top, the idealized polar update Q=UV⊤Q=UV^\top gives every singular direction the same weight. We refer to this as the flat profile. Several recent optimizers replace this flat profile with fine-grained spectral maps that give each direction its own gain. We ask how much of this spectral detail a Muon update needs. Our spectral diagnostics show that approximately 9494--97%97\% of measured singular modes lie below an estimated noise edge, yet collectively align positively with a reference gradient. We introduce BulkBoost, a two-band spectral reweighting framework with fixed-rank and noise-calibrated variants. The latter uses split-minibatch gradient differences to calibrate a Marchenko--Pastur reference edge for Muon's Nesterov input, separating the bulk below the edge from the spikes above it. Both variants increase the bulk's relative weight through one shared gain while preserving the Frobenius norm of each matrix's unreweighted direction. For a fixed partition, our theory gives the first-order condition under which moving weight toward the bulk lowers the loss. It also quantifies the fraction of the maximal first-order improvement rate, over all per-mode reallocations, that two bands can capture. Across 30 continued-pretraining settings spanning Pythia-14M to 410M and six corpora, two-band reweighting is competitive with the fine-grained power-law profile of Freon and outperforms Spectra. Measured against Muon's flat profile, Freon reduces final loss by 0.022%0.022\% of the pre-adaptation loss on average, whereas the two-band variants achieve reductions of 0.0730.073--0.147%0.147\%. These observations suggest that useful departures from the flat profile are surprisingly low-dimensional: a single bulk-to-spike gain captures at least as much benefit as the fine-grained spectral profiles.
Oct 5, 2026

DGA-Muon: Decoupled Geometry-Aligned Adaptive Scaling for Muon

While NorMuon has achieved strong empirical performance in pretraining, its underlying adaptive mechanism remains largely heuristic and poorly understood. In this work, we provide the first systematic theoretical analysis of NorMuon's adaptivity, revealing that it primarily arises from orthogonalization-induced geometry rather than genuine optimization dynamics, serving to offset the resulting geometric non-uniformity. Under exact orthogonalization, the adaptive scaling factors degenerate into a single global scalar for square and wide matrices, while for tall matrices their variation results from unevenly distributed row energy after orthogonalization. Under approximate orthogonalization, the orthogonality residuals introduce additional variation into the scaling, giving rise to a counterintuitive Orthogonalization--Adaptivity Paradox: more accurate orthogonalization weakens adaptivity. We further show that NorMuon's rigid row-wise scaling is geometrically misaligned with the one-sided orthogonal structure of tall matrices by distorting column orthogonality. Motivated by these limitations, we propose two core design principles that a desirable adaptive mechanism for Muon should satisfy. First, adaptive scaling should be decoupled from orthogonalization, with the scaling factors computed directly from raw gradients. Second, adaptive scaling should be aligned with the shape-dependent orthogonal structure of the polar factor, using row-wise scaling for wide matrices and column-wise scaling for tall matrices. We prove that this geometry-aligned scaling preserves the orthogonal structure of the update. By incorporating several other techniques, we obtain Decoupled Geometry-Aligned Muon (DGA-Muon). We establish convergence guarantees for DGA-Muon and empirically validate both our theoretical characterization of NorMuon's scaling degeneration and the superiority of DGA-Muon.
Oct 5, 2026

Muon Is Theoretically Wrong For Convolutions, But Empirically Effective

Muon, an optimizer known for its efficiency, has a clear interpretation for matrix-valued updates, but convolutional kernels are stored as four-dimensional tensors. Standard implementations reshape these tensors into matrices, a shortcut which breaks the theoretical understanding behind Muon. To investigate this, we formalize the corresponding optimization objective directly in convolutional operator geometry and introduce Convolutional Newton-Schulz (Conv-NS), which approximates the polar factor in this geometry while preserving kernel support. When applied in fast training experiments, Conv-NS and reshape-based Muon are both computationally efficient and achieve comparable accuracy on CIFAR-10 and ImageNet classification tasks. However, as one could expect a theoretically aligned Conv-NS to outperform reshape-based Muon, we investigate this mismatch between practice and theoretical understanding, with the hypothesis that exact convolutional orthogonalization may overconstrain updates. These findings highlight Muon's strong practical performance while opening directions for its further development on convolutions. Our code is publicly available at github conv-muon.
Oct 2, 2026

Training-Loss Guarantees for Muon with Finite-Step Newton--Schulz Orthogonalization

Existing convergence analyses of Muon either assume exact orthogonalization or analyze classical Newton--Schulz polynomials, and guarantee only stationarity, so it is unresolved what Muon's five tuned Newton--Schulz steps preserve and whether that suffices to reach a prescribed neural-network training loss. We establish a finite-time training guarantee that accounts for both momentum accumulation before orthogonalization and the tuned finite-step update. For full-batch training of a sufficiently wide two-layer ReLU network with fixed random output weights and a positive-definite limiting neural tangent kernel, we prove that Muon reaches any target empirical squared loss ε>0\varepsilon>0 with high probability over initialization. For every momentum parameter μ∈[0,1)μ\in[0,1), a target-dependent constant learning rate proportional to (1−μ)ε(1-μ)\sqrt{\varepsilon} yields a hitting-time bound of O((1−μ)−1ε−1/2)O((1-μ)^{-1}\varepsilon^{-1/2}), with other problem parameters fixed. The sufficient width is independent of both target accuracy and momentum. The analysis shows that the tuned Newton--Schulz map preserves alignment with the momentum buffer while bounding the update's spectral norm. Control of gradient variation near initialization transfers this alignment to the current gradient, ensuring descent until the target is reached without requiring exact orthogonalization. Numerical experiments support these mechanisms at widths below the sufficient theoretical threshold: gradient-update alignment remains above the analytical reference, and all 30 runs across six widths and five student initializations on a fixed teacher-student dataset reach the target loss while maintaining kernel positivity.
Oct 2, 2026

Muon Learns Facts Better: Understanding the Role of Spectral Orthogonalization

The Muon optimizer applies spectral orthogonalization to matrix-valued updates and has shown strong performance in large-scale neural network training, yet the mechanisms of this transformation in feature learning remain poorly understood. In this work, we investigate this question through a tractable factual-recall model, where a fact maps each subject-relation pair to an answer, and a linear transformer learns the subject- and relation-dependent information required to recover this mapping. The transformer is optimized with gradient flow (GF), spectral GF, or Sign GF, which are continuous-time limits of gradient descent, Muon, and Adam, respectively. Prior studies (Nichani et al., 2025) have shown that when the number of subjects exceeds the number of relations, GF learns relation-dependent information before subject-dependent information, producing a feature-separation phase during training. We characterize this separation with the learning times when the subject- and relation-dependent components of the prediction reach a target accuracy. With SS subjects and RR relations, GF has a learning-time ratio of Θ~(S/R)\widetildeΘ(\sqrt{S/R}), whereas Spectral GF reduces this ratio to Θ~(1)\widetildeΘ(1). In addition, for fixed SS and RR, the subject- and relation-dependent errors decay as 1/(Tlog⁡T)1/(T\log T) in training time TT under GF, but as exp⁡(−poly(T))\exp(-\mathrm{poly}(T)) under spectral GF. Finally, we show that GF and spectral GF are equivariant under orthogonal transformations of the token embeddings, whereas Sign GF is not: Different orthonormal embeddings can potentially produce no feature separation, a large feature-separation phase, or even a reversed learning order. These results provide a mechanistic view of how spectral orthogonalization can fundamentally reshape feature-learning dynamics.
Oct 2, 2026

Inner Momentum for Differentially Private Muon

Differentially private training clips each per-example gradient before adding noise. This clipping is radial for each example, yet unequal clipping factors can distort the relative singular-vector geometry of their average. Muon is particularly exposed to this effect, since its update is an approximate polar factor UV^T that depends only on the singular vectors that clipping can shift. To curb this degradation, we propose averaging each sampled example's Muon gradient over the current model and a short history of recent models before clipping. The clipped batch matrix then separates into a common rescaling and a covariance residual R between sampled gradients and clipping values, with ||R||_F <= sigma_lambda sigma_G, bounding the clipping-induced distortion directly. We further show that a finite Newton-Schulz iteration preserves the polar factor of its input under these spectral conditions, confirming that our correction survives orthogonalization. In private GPT-2 fine-tuning on E2E and DART at epsilon in {1, 2, 4, 8}, DP-Muon-IM improves BLEU and ROUGE-L over DP-Muon in every seed-matched comparison, and non-private diagnostics show 2-4% lower pre-noise polar error.
Oct 2, 2026

MuonIO: Principled Norm-Aware Descent for Embedding Tables and Language Model Heads

The Muon optimizer derives its update rule for hidden linear layers by solving a local linearization of the loss penalized by the spectral norm, motivated by an RMS-stability argument for dense linear layers. Standard Muon implementations, however, exclude the input (embedding table) and output (language model head) layers from this principled treatment, for which they use AdamW instead. We present MuonIO, a single Muon-style update for both of these layers. For the language model head L∈RV×d\mathbf{L} \in \mathbb{R}^{V \times d}, we motivate the use of the 2→∞2\to\infty operator norm, due to the Lipschitz continuity of the softmax output geometry, while for the embedding table E∈Rd×V\mathbf{E} \in \mathbb{R}^{d \times V}, we draw on the 1→21 \to 2 operator norm, based on the one-hot input geometry identified by Bernstein & Newhouse (2025). The identity ∥L∥2→∞=∥L⊤∥1→2\lVert\mathbf{L}\rVert_{2\to\infty}=\lVert\mathbf{L}^\top\rVert_{1\to2} then puts both matrices in the same vocabulary-oriented geometry: MuonIO applies a single normalized-vector rule, which appears as column normalization for E\mathbf{E} and row normalization for L\mathbf{L}. Empirical evaluations demonstrate the effectiveness of our approach, with MuonIO reducing I/O optimizer state memory by 50% and I/O update FLOPs by ∼\sim46% compared to Muon for 1B LLaMA pretraining on C4, while also improving validation perplexity.
Oct 1, 2026

AF-Muon: An AdamW-Free Muon Optimizer for Tied-Embedding Models

Muon improves large-scale training by applying a spectral-norm steepest-descent update to matrix parameters, but practical models also contain parameter blocks that do not fit dense-matrix geometry. One important case is the tied vocabulary table, which appears in language models and other token generators and can receive multiple structurally different gradient sources, from sparse input lookups to dense output-classifier updates. In the reference recipe these blocks are handed to an auxiliary AdamW optimizer, which restores second-moment state and updates the aliased table as a generic tensor. We propose AF-Muon, an AdamW-free extension of Muon that keeps the Muon matrix update for hidden weight matrices while using a support-aware finite-cap linear minimization oracle for tied vocabulary tables and an RMS-normalized update for one-dimensional auxiliary parameters. AF-Muon therefore trains every parameter class with a single first-moment buffer and no second-moment state, saving around 20% optimizer-state memory relative to Hybrid Muon in our benchmark. Across nine tied-token settings - decoder-only language models from 124M to 1B parameters, a fully shared T5-style encoder-decoder, and ImageGPT-style image-token, protein, and sparse-MoE variants, spanning text, image, and protein-sequence data - AF-Muon improves mean validation loss and perplexity over both Hybrid Muon and a SCION-style Sign endpoint. Long-horizon runs and hyperparameter sensitivity studies confirm the gain is robust, and identical-momentum diagnostics attribute it to the finite cap, which preserves more within-row magnitude than Sign while bounding the coordinate concentration of row-RMS. These results identify tied vocabulary tables as a distinct optimizer geometry and yield a robust AdamW-free Muon variant across models, modalities, and architectures, with about 1% step-time overhead in matched training.
Sep 30, 2026

Convergence of Practical Muon with Finite Newton-Schulz Iterations and Nesterov Momentum

Practical Muon maintains momentum and performs a small, fixed number of Newton--Schulz iterations separately for each parameter matrix, often with a Nesterov correction. We analyze these layer-wise finite-step updates jointly on a coupled nonconvex objective, rather than replacing them by exact polar factors or one global orthogonalization. Under gradient-dependent (L0,L1,q)(\mathcal L_0,\mathcal L_1,q)-smoothness and conditionally unbiased stochastic gradients with bounded layer-wise variance, we establish an O(T−1/4)\mathcal O(T^{-1/4}) bound on the expected average Frobenius gradient norm. The analysis retains the Nesterov recursion and requires neither bounded stochastic gradients, symmetric noise, nor a uniform positive lower bound on the nonzero output singular values. Its constants contain no explicit matrix-dimension or rank factors when the number of blocks and problem constants are fixed. The proof follows a descent inequality and a decomposition of the momentum tracking error into initialization, noise, and drift. For the original five-step quintic, we verify the required scalar-map bounds analytically; the result also allows step-dependent coefficients satisfying the same bounds. A complementary nuclear-norm result quantifies rank dependence under a stronger spectral condition. The vanishing rate uses coupled learning-rate and momentum schedules, including the standard single-coefficient Nesterov rule.
Sep 30, 2026

The Row Normalization Puzzle in Muon

This paper examines how row-wise renormalization affects Muon, focusing on the gap between NorMuon's worst-case guarantees and its practical performance (Li et al.). Despite its growing adoption and promising performance in large language model (LLM) pretraining, NorMuon's worst-case guarantees remain poorly understood. One fundamental question is: Does row normalization yield provable convergence gains, potentially through its interaction with approximate polar computation and exponential moving-average momentum? Our results show that row normalization introduces a dimension-dependent factor in the worst-case iteration complexity under the operator-norm geometry, which persists even with exact polar computation and any fixed momentum parameters. Indeed, we establish an algorithm-dependent lower bound and a matching upper bound in deterministic settings, and extend our upper bound analysis to stochastic settings. Both upper-bound analyses allow approximate polar computation. Experiments show that NorMuon is slower than Muon on synthetic problems inspired by our worst-case construction, yet outperforms Muon in LLM pretraining. These findings sharpen the puzzle of why row normalization helps in practice and complement the recent findings of Dewulf et al.
Sep 29, 2026

Physical Muon: Orthogonalization as an Equilibrium Computation

Physical neural networks and analog in-memory computing could reduce the energy cost of neural network training. Realizing this potential, however, requires optimizers that combine effective learning with physical implementability. SGD fits local analog updates but struggles on transformers, while Adam family is unstable against analog bias. Muon offers strong training performance, but its Newton--Schulz orthogonalization relies on dense matrix-matrix products. To address this obstacle, we introduce Physical Muon, which computes the orthogonalization as the equilibrium of a continuous-time flow. Random probes approximate the flow using matrix-vector products, reciprocal reads, and local rank-1 writes. To test whether this replacement preserves training performance, we evaluate it on a 10.95M-parameter transformer. The dense flow's mean validation cross-entropy is 0.0085 above Newton--Schulz across nine seeds per method; the probe implementation is 0.0188 above the control across two seeds. Circuit simulations further reproduce the flow dynamics and yield comparable training behavior.
Sep 29, 2026

Second-Moment Stochastic Approximation Methods

Classical stochastic approximation methods rely on estimators of the first moment (mean) of a random regression function. We study methods that employ estimators of both the first and the second moments, which include modern deep-learning optimizers such as Adam and Muon as special cases. We derive second-moment stochastic approximation methods through the lens of optimal preconditioning for solving matrix equations, and develop a two-stage framework for their convergence analysis. The first stage focuses on the analysis of conceptual (impractical) methods that rely on the exact first and second moments. In the second stage, we replace the exact moments with their respective estimators, and invoke Dvoretzky's theorem to show that the resulting practical methods converge almost surely to a neighborhood of the target solution. The size of the neighborhood depends on the biases and variances of the first- and second-moment estimators. We derive concrete bounds for Muon and a spectral variant of Adam that determine the radius of their neighborhood of convergence.
Sep 28, 2026

Can Representation Learning Decouple from Loss Minimization? Polar Updates Have an Answer

Does representation learning stop when the training loss stops improving? We study this question for matrix Muon, whose polar-normalised updates have a step length set by the gradient's rank rather than its norm. Near the edge of stability, full-batch Muon on teacher-student problems enters approximately period-2 loss oscillations that persist for thousands of steps: the cycle-mean loss stays flat or rises, yet the weights keep moving and the learned features continue to align with the teacher subspace. For linear teacher-student learning toys, we derive explicit cycle and alignment formulas and conditional plateau and decay bounds. For a population mean-field ReLU model, we prove that, under stated dimension, initialisation and small-head conditions, the leading eigenspace of the average gradient outer product (AGOP) recovers the teacher subspace exactly during a loss plateau, before the loss later drops. In all 33 ReLU, GELU and SiLU teacher configurations we study, direction-only alignment metrics show the student AGOP aligned with, or still aligning to, the teacher subspace during the period-2 oscillations; projected head refitting on selected configurations shows that the learned directions are useful for prediction, and further measurements distinguish AGOP alignment from weight-mass concentration. In deep residual ReLU students, freezing the downstream layers while the first layer trains with full-batch exact polar updates recreates a nearly flat cycle-mean loss with improving input-AGOP alignment; freezing and unfreezing switch between this plateau and loss decrease, and the effect is sensitive to momentum and to the choice of orthogonaliser.
Sep 28, 2026

Muon Sublates the Edge of Stability in LLM Pretraining

Muon is increasingly used for language-model pretraining, yet its large-step dynamics are not captured by the classical edge-of-stability (EoS) picture of gradient descent (GD). In GD, loss neutrality, equal-magnitude update reversal, and marginal stability meet at a single learning-rate-dependent edge. We show that Muon breaks this coupling. For stochastic no-momentum Muon, we derive a coherence-corrected conditional loss-neutral boundary 2ρb/η2ρ_b/η, while temporal alignment follows a separate geometry. Controlled experiments show that loss balance and temporal alignment respond differently to learning rate and batch size. Across our language model experiments, the 130M Llama-like LLM runs exhibit loss-boundary tracking with weak negative alignment, whereas the studied 1B LLM configuration shows stronger partial cancellation; in both settings, directions remain far from coherent reversal while training continues to improve. These results support a split EoS picture for Muon: a stochastic loss-neutral edge survives, but it is not accompanied by a universal temporal-direction signature. The source code for reproducing the experiments can be found in https://github.com/cyzebra/Muon-Sublates-the-Edge-of-Stability-in-LLM-Pretraining
Sep 27, 2026

Orthogonal Witness Control for Muon Optimization via Sigmoid Spectral Reshaping

Matrix-valued optimizers such as Muon exploit the spectral structure of neural network updates through Newton--Schulz orthogonalization, but their near-flattening of the singular spectrum discards relative magnitude information across gradient modes. We introduce Soren (Spectral Orthogonal Reshaping), a matrix-valued optimizer that preserves the singular subspaces of the gradient while applying a bounded, monotone sigmoid transformation to its singular values. This smoothly compresses dominant modes without fully flattening the spectrum. We interpret Soren as a positive-definite preconditioned gradient method and establish convergence guarantees under relative smoothness and metric Polyak--Łojasiewicz geometry. To avoid explicit singular value decomposition, we further develop a finite-depth Soft Newton--Schulz (SNS) polynomial realization of the sigmoid spectral map and characterize how its spectral approximation affects the induced convergence geometry. Experiments across LLM pre-training, supervised fine-tuning, and direct preference optimization demonstrate the effectiveness and robustness of Soren against established optimizers.
Sep 27, 2026

Convergence of Practical Muon

Muon is emerging as a promising alternative to AdamW for large-scale neural network training, yet theoretical understanding of its practical implementation remains incomplete, as existing analyses often simplify or omit two key components: (i) practical Newton--Schulz iterations with empirically tuned polynomial coefficients (3.4445,−4.7750,2.0315)(3.4445,-4.7750,2.0315); and (ii) decoupled weight decay for regularization. In this paper, we provide an optimization interpretation and establish convergence for practical Muon, jointly accounting for both components. Specifically, we interpret practical Muon as right-preconditioned optimization of the original loss with a dynamic weighted ℓ2\ell_2 regularizer that vanishes as stationarity is approached, so that the optimization target remains the original objective. We then establish, to our best knowledge, the first convergence guarantee for practical Muon in the stochastic nonconvex setting, with an O(T−1/4)\mathcal{O}(T^{-1/4}) convergence rate in terms of the expected Frobenius norm of the gradient, improving the dimension dependence of the best known AdamW's convergence rate by a factor of d\sqrt{d}, where TT is the iteration horizon and dd is the parameter dimension. Experiments further support the theoretical convergence results.
Sep 27, 2026

Cost-free Spectral Estimation for Adaptive Newton--Schulz in Matrix Optimizers

Matrix optimizers such as Muon transform each momentum matrix through an approximate orthogonalization, typically implemented by a small number of Newton-Schulz matrix multiplications. The quality and cost of this approximation depend strongly on the singular-value spectrum of its input, yet existing implementations use the same fixed polynomial routine for every layer and throughout training. We show that this uniform treatment is unnecessary: the computations in the Newton-Schulz method already reveal enough information to make the method adaptive. The Gram matrices formed inside Newton-Schulz iterations yield spectral moments through inexpensive scalar reductions, requiring no additional matrix multiplications. From these moments, we recover an estimate of the empirical singular-value distribution and use it to select a polynomial routine specialized to the current matrix. This turns Newton--Schulz orthogonalization into a spectrum-adaptive procedure that responds to differences across both layers and training time. On saved momentum matrices, spectral estimation substantially reduces orthogonalization error at a fixed iteration budget or reaches the same accuracy with fewer iterations, and in GPT pretraining up to 1B parameters it lowers the validation loss of two matrix optimizers. Our results suggest that matrix-function operations inside optimizers need not be designed for a conservative worst-case spectrum: they can cheaply measure the spectrum they are already processing and specialize computations accordingly.
Sep 26, 2026

Muon Under Gradient Noise and the Limits of Orthogonalization Near Optima

Muon replaces the momentum buffer of each weight matrix by its orthogonal polar factor. We ask what this orthogonalization does near an optimum, where minibatch noise dominates the gradient. Under a Gaussian noise model, the expected Muon update becomes a scaled gradient step, so a linear method with a suitably matched learning rate reproduces Muon's first-order mean response. The stochastic update is a different matter: after the response is matched, Muon retains a nonlinear residual that is uncorrelated with the input noise and contributes additional covariance. A Hermite expansion shows how momentum acts on this residual. Its higher-order components decorrelate faster than the linear component, so momentum suppresses the residual's accumulated covariance relative to the linear part, but never eliminates it. In a local quadratic surrogate that evaluates the residual on the stationary noise buffer, the residual adds stationary covariance and raises the stationary loss floor at every stable step size, while leaving the contraction dynamics unchanged. Simulations of the full nonlinear recursion on quadratics and measurements on frozen transformer gradients support each step of this picture. Together, the results make a theoretical case for replacing orthogonalization by response-matched momentum SGD once optimization becomes noise-dominated.
Sep 24, 2026

Convergence guarantees for Muon: New parameter regimes and generalizations

In this paper, we establish the first asymptotic convergence guarantees for the Muon algorithm through a more accurate proxy for the Newton-Schultz iteration than the typical matrix sign function. We prove that, for appropriate choices of hyperparameters, the iterates satisfy lim⁡k→∞∥∇f(xk)∥=0\lim_{k\to\infty}\|\nabla f(x_k)\|=0, and, under a global Polyak-Łojasiewicz condition, that the sequence of function values converges linearly. The key insight is that the regularization, implicit in Muon's Newton-Schulz implementation, induces a bounded preconditioner, exposing Muon as a preconditioned Polyak heavy-ball method and enabling a classical Lyapunov analysis. This observation naturally motivates applying the same preconditioning structure to the Nesterov gradient evaluation. We formalize this idea by introducing Muesterov, a Nesterov-based variant of Muon, and prove that it enjoys the same convergence guarantees, extending the theoretical framework beyond the heavy-ball setting. Numerical experiments on a scalar cross-entropy problem corroborate the theory and illuminate the joint role of the learning rate and the Newton-Schulz regularizer in controlling convergence. Preliminary numerical simulations training the nanoGPT dataset provide intuition regarding the relevance of the observations in this paper to practical applications.
Sep 21, 2026

Muon Can Outperform Dedicated Continual Learning Methods

Continual learning with Low-Rank Adapters (LoRA) typically mitigates forgetting by penalizing the overlap between a new update and the accumulated past weights, which discourages certain update directions without controlling how an update distributes its energy over the ones that remain. We ask whether that restriction has to be task-aware, or whether a generic one supplied by the optimizer is enough. We train a plain incremental LoRA (IncLoRA) with Muon, which orthogonalizes each update, and compare it against O-LoRA and ELLA over five seeds and three task orders on the Standard CL Benchmark and three seeds on TRACE. IncLoRA+Muon reaches the accuracy band of the dedicated methods on Standard CL and improves on every AdamW configuration on TRACE. One update-constraining mechanism is enough, whether it comes from the loss or from the optimizer; on Standard CL a second one does not help, and for the most restrictive method it costs 8.4 points of accuracy and the plasticity to fit each task. What separates the two optimizers is not the size of the update, which under Muon is 0.91 to 2.06 times that under AdamW, but how it is distributed. AdamW confines it to between 1.4 and 1.8 effective singular directions, Muon spreads it over 7.0, and the two do not overlap in any tracked run. Part of the advantage usually attributed to dedicated CL methods may therefore be explained by the geometry of the optimizer's updates.
Sep 15, 2026

Derivative-Free Structured Updates for Muon

Muon updates matrix-valued neural-network parameters by orthogonalizing a gradient-based momentum matrix. Its reliance on derivatives limits its use when gradients are unavailable or unreliable. We develop a derivative-free framework that constructs Muon-style updates from structured finite differences. Four variants are considered: full entrywise recovery, random low-rank surrogates, basis-aligned rank-one probing, and direct structured search. Exhaustive basis-aligned probing is equivalent, up to positive scaling before ideal polar orthogonalization, to coordinate finite differences. Matrix-regression experiments show that random rank-one probing can reduce the number of function evaluations substantially, at the cost of less accurate updates. Controlled noisy-gradient experiments on regression and a neural network illustrate when accurate function values can compensate for an unreliable gradient oracle. A small CartPole study further examines orthogonal rank-one probes under a fixed episode budget. These results support structured probing as a practical option for selected black-box problems; they do not establish a general convergence guarantee or an advantage over accurate, inexpensive gradients.
Sep 12, 2026

Nonsmooth Optimization via Orthogonalized Momentum

Modern real application problems involve matrix-valued parameters, yet conventional optimizers treat them as vectors, thereby motivating matrix-aware methods that exploit input-output geometry, such as Muon which orthogonalizes the momentum matrices before parameter updates. Its empirical success raises a conceptual question: can orthogonalized momentum remain effective beyond smooth optimization? This paper studies this question for locally Lipschitz functions using a generalized derivative framework compatible with backpropagation. Our first contribution is to identify a key limitation: for every fixed momentum factor β∈[0,1)β\in[0,1), Muon can fail to approach the global optimal solution of a convex Lipschitz objective from almost every initialization, when step sizes adapt to the full gradient history. The failure can occur even along bounded iterates. Our example is inspired by the one of Parshakova et al. which only covers β∈[0,12)β\in[0,\frac{1}{2}). Then, we show that the obstruction lies in fixed momentum rather than orthogonalization. Indeed, when the momentum factor is adaptive and approaches 1 together with a vanishing step size, Muon recovers asymptotic convergence for nonconvex nonsmooth optimization under the boundedness and regularity conditions. Moreover, we propose MAGD, which combines orthogonalized momentum with gradient, weighted based on their relative progress. MAGD retains asymptotic convergence in nonconvex settings and achieves an O(min⁡{m,n}ε−2)O(\min\{m,n\}ε^{-2}) rate in convex settings. A lower bound shows the optimal dimension dependence. Experiments on synthetic problems, image classification, and LLM pretraining show MAGD is a simple and practical alternative to Muon. Together, our results characterize when orthogonalized momentum fails without smoothness and how it can be made reliable and we hope that the analysis may be useful more broadly.
Sep 10, 2026

Musec: MomentUm SpEctral Clipping for Stable Muon-type Training

Muon has emerged as a highly effective optimizer for large language model training, often achieving superior convergence and performance compared with the widely adopted Adam and AdamW optimizers. Nevertheless, Muon is prone to training instability due to its spectral flattening, manifested by loss spikes and unbounded growth of model weights. Existing approaches primarily rely on weight or attention-logit clipping, which require architecture-specific modifications and do not directly address instability across all model components. We propose MomentUm SpEctral Clipping (Musec), which replaces Muon's spectral flattening with spectral clipping: rather than setting all singular values of the momentum matrix to approximately one, Musec clips singular values that exceed a threshold while preserving the underlying spectral structure of the momentum. Our strategy provides an optimizer-level, architecture-agnostic mechanism for stabilizing Muon training. Theoretically, we establish convergence guarantees for Musec in nonconvex nonsmooth stochastic optimization. Practically, we develop Soft Musec, an efficient implementation that uses a smooth spectral saturation function approximated by coupled Newton-Schulz iterations. Empirically, Soft Musec consistently improves training stability over existing Muon variants across a wide range of learning rates and model sizes, remaining stable in settings where existing Muon variants diverge while matching their performance under well-tuned configurations.
Sep 9, 2026

Muon-C: Operator-Aligned Muon for Convolutional Kernels

Muon replaces matrix momentum with an approximately orthogonal polar direction, but its geometry depends on the matrix representation. For convolution, standard unfolding describes a local patch map rather than the convolution operator. We introduce Muon-C, an operator-aligned optimizer that represents kernel momentum as frequency-wise channel-transfer matrices, polarizes these blocks independently, and uses a critical Fourier grid to return updates exactly to the original finite kernel support. We show that the new geometry arises from combining the block partition and Fourier coordinates. The exact-polar direction is a linear minimization oracle under the critically sampled convolution norm. Its worst-case guarantee relative to the continuous convolution-operator norm is never weaker than unfolding and is strictly stronger for 3×33\times3 kernels. On CIFAR-10 flow matching with matched applied-update RMS, Muon-C reaches 9.87 FID at 40k iterations, compared with 22.26 for unfolded Muon and 51.31 for Adam. It reaches their final quality using 0.62×0.62\times and 0.64×0.64\times their model FLOPs, respectively. Under equal tuning budgets, Muon-C achieves 3.42 FID. Gains persist across data scales and transfer to classification across convolutional architectures.
Sep 7, 2026

Beyond the Matrix Sign: Quadratic Spectral Descent

Muon emerges as a strong competitor of the AdamW for LLM pretraining, because the matrix-wise update it employs can potentially incur smaller second-order penalty than the once dominating AdamW, which performs coordinate-wise update. However, the spectral flattening procedure in Muon is quite debatable since it discards the spectral amplitude information totally. To seek for better spectral allocation (and the associated spectral subspace), we propose to solve the quadratic model of loss function under the spectral norm constraint directly (i.e., in a genuinely Newtonian way) and thus obtaining the Quadratic Spectral Descent (QSD) algorithm. In contrast, many existing curvature-aware methods either exploit the second-order information in an implicit way by changing the weight update geometry (such as Mousse, FISMO) or rely on strong assumptions (such as the weight displacement isotropy assumption in Newton-Muon). QSD's potential advantage over these methods is best illustrated in the isotropic curvature scenario, where Mousse, FISMO and Newton-Muon all reduce to Muon while the spectral allocation in QSD is still non-flat (since the spectral allocation in QSD depends on the gradient to curvature ratio). Meanwhile, to control the complexity of QSD, we employ inversion-free K-FAC and online Frank-Wolfe update which is essentially a matrix sign operator. Overall, the complexity increase can be rather mild. Experiments on GPT pre-training show that QSD consistently improves validation loss over Muon and recent Muon variants, while achieving up to an 8.49%8.49\% wall-clock speedup at matched validation loss.
Sep 5, 2026

FedSubMuon: Communication-Efficient Federated LLM Fine-Tuning via Structured Subspace Muon

Federated fine-tuning adapts large language models (LLMs) to decentralized client data, but its scalability in cross-device training is often limited by the high communication cost. Muon is an optimizer that improves optimization performance by orthogonalizing momentum for matrix-valued parameters. Existing federated Muon methods demonstrate the benefit of matrix-aware optimization in federated learning, but still require transmitting full layer-size updates and optimizer state. A natural way to reduce communication is to directly apply Muon to LoRA factors, but this changes the optimized object and weakens Muon's matrix-aware update geometry. We propose FedSubMuon, a communication-efficient federated Muon fine-tuning method that optimizes compact coefficient matrices within shared structured subspaces. This design keeps Muon on a single matrix-valued trainable object, while reducing the client upload to compact coefficient matrices. We further introduce FedSubMuon-GT, an accuracy-oriented extension that uses projected gradients to adapt tracked subspace bases toward task-relevant gradient directions. Experiments on instruction tuning and mathematical reasoning show that FedSubMuon-GT achieves the best overall accuracy on four of five dataset-model pairs, while FedSubMuon performs best under all matched communication budgets. On Dolly-15K, the closest communication baseline requires 5.5 times and 1.4 times more total communication on Llama-1B and Qwen-4B, respectively.
Sep 4, 2026

Optimizer Memory Schedules for Outscaling the Overtraining Axis

We investigate how optimizers scale across the overtraining axis and show that relative optimizer performance and optimal hyperparameters change substantially with training horizon. In particular, we study how matrix-preconditioned methods (Muon and SOAP) and a momentum-scheduled method (ADANA) scale relative to AdamW. We compare these four optimizers across models from 51M to 253M parameters and overtraining (OT) factors from 1x to 256x, sweeping the base learning rate at every setting. The preferred learning rate schedule can reverse across the overtraining axis, the best weight decay coefficient scales approximately as sqrt(OT), and longer horizons generally favor longer fixed memory. ADANA's scaling advantage over AdamW persists after tuning AdamW's fixed memory separately at each horizon. Log-time weight decay and momentum cooldown provide substantial gains for ADANA that compound as training increases. With this treatment, ADANA outscales AdamW with an exponent advantage close to that predicted by DANA theory on power-law random features. Muon and SOAP instead provide roughly constant token-efficiency advantages over AdamW across most of the measured range, although SOAP may gain further at the highest overtraining factors. ADANA begins behind both matrix-preconditioned optimizers but closes the gaps as training increases, surpassing Muon and becoming competitive with SOAP at our highest OT factors. These results establish training horizon as an essential axis for optimizer evaluation and design.
Aug 27, 2026

When Muon Meets Task Interference: A Spectral Perspective on Continual Learning and Model Merging

Continual learning (CL) and model merging (MM) both aim to obtain a single model that performs well across multiple tasks, challenged respectively by catastrophic forgetting and weight-disentanglement error. In the literature, these difficulties are merely treated separately and mitigated through a variety of solutions, while the geometry induced by the base optimizer is treated as an implementation detail. In this work, we show that the two difficulties are in fact two instances of the same phenomenon: a parameter update useful for one task shifts the model's outputs on another. We formalize this shared phenomenon as task interference and reduce it to a common layer-wise Frobenius inner product ⟨ΔWℓ,Jℓ(x)⟩F\langle ΔW_\ell, J_\ell(x)\rangle_F. This quantity, in turn, is utilized to expose the role of the optimizer. We theoretically derive an upper bound that isolates the spectral norm ∥ΔWℓ∥2\|ΔW_\ell\|_2 as an optimizer-controllable factor of task interference, and a per-mode analysis shows that this bound tracks the dominant part of the empirical interference. Specifically, we then identify the recent Muon optimizer as a mechanism that regulates this factor by construction. Our work reveals that its elegant control on spectral norm tightens the interference bound for both CL and MM, positioning Muon as a principled optimizer-centric approach complementary to existing solutions. Our theoretcal analysis is well validated by experimental results. Replacing the AdamW optimizer with Muon improves accuracy by up to +5.02 points on the eight-task model-merging benchmark across three CLIP backbones. For continual learning, Muon also delivers uniformly positive gains across ten class-incremental protocols, three task-incremental protocols, and the 11-task MTIL benchmark.
Aug 26, 2026

Muon with Finite Newton-Schulz: The Smoothing Benefit in Nonsmooth Nonconvex Optimization

Muon has emerged as a strong optimizer for the matrix-valued parameters in large language model pretraining, approximately orthogonalizing its momentum with a few Newton-Schulz iterations. Existing theory either replaces this iteration with the exact polar factor it approximates, or treats its finite depth as an approximation error, and thus the iteration Muon actually runs can only hurt the guarantees. We show that finite Newton-Schulz can instead be beneficial for nonsmooth nonconvex optimization. To this end, we analyze Muon through the online-to-nonconvex conversion, which views the update rule as an online learner and converts its regret bound into a stationarity guarantee. The finite Newton-Schulz iteration smooths the discontinuous polar map into a Lipschitz map of the singular values, and Muon with finite Newton-Schulz can be regarded as an online learner with a smoothed spectral potential. This smoothing is exactly what the conversion needs: we prove that a Newton-Schulz depth growing only logarithmically in the target accuracy suffices for convergence to stationary points in nonsmooth nonconvex optimization, whereas Muon with the exact-polar update may fail to converge. The resulting sample complexity bounds match the best-known guarantees for nonsmooth nonconvex optimization and are optimal for smooth nonconvex optimization up to problem-dependent factors. The argument extends beyond Newton-Schulz to general spectral maps with the same smoothing property.
Aug 26, 2026

Spectral Allocation: Why Muon Outperforms Adam, and How to Improve Muon

Orthogonal optimisers such as Muon can substantially accelerate large language model pretraining relative to Adam, yet the mechanism remains incompletely understood. We investigate this through an out-of-sample spectral probing analysis of Transformer loss landscapes. At checkpoints along real training trajectories, we decompose each momentum buffer into its singular directions and estimate the loss-optimal step size along each direction on held-out data. The resulting spectral profile is anisotropic yet stable across batches and training stages, and consistent across the optimisers and model scales: a volatile head operating at the Edge-of-Stability supports a much smaller step size than the tolerant bulk, which permits substantially larger steps. This profile provides a unified spectral allocation account of why Muon outperforms Adam, which outperforms SGD. It also exposes a limitation of Muon's uniform scaling: it still underutilises the bulk. Guided by this finding, we introduce Spectral-Aware Muon (SAMuon), which holds the head at the Muon scale and amplifies the bulk using a static spectral prior. We provide two variants: the complete SAMuon follows the measured profile using a low-rank randomised SVD and the simplified SAMuon-lite uses a two-level approximation via rank-one power iteration. Neither method adds persistent optimiser state or notable extra FLOPs beyond Muon at scale, and the idealised exact-whitening versions of both retain Muon's asymptotic convergence rate under standard assumptions. Across "modded-nanogpt" models from 124M to 1B parameters, both variants outperform tuned AdamW and Muon (Scion implementation) baselines in all evaluated model-scale and batch-size configurations. SAMuon requires 13.3% to 24.0% fewer training tokens to reach the same validation loss as Muon, while SAMuon-lite retains most of this gain with near-zero wall-clock overhead.