Dead-Direction Conditioners provide gauge-equivariant preconditioning by conditioning optimizer state on symmetry orbits, yielding improved resistance to over-training collapse and higher detection of dead directions compared to AdamW and Muon.
hub Canonical reference
Training Deep Learning Models with Norm-Constrained LMOs
Canonical reference. 100% of citing Pith papers cite this work as background.
36
Pith papers citing it
Background
100% of classified citations
abstract
In this work, we study optimization methods that leverage the linear minimization oracle (LMO) over a norm-ball. We propose a new stochastic family of algorithms that uses the LMO to adapt to the geometry of the problem and, perhaps surprisingly, show that they can be applied to unconstrained problems. The resulting update rule unifies several existing optimization methods under a single framework. Furthermore, we propose an explicit choice of norm for deep architectures, which, as a side benefit, leads to the transferability of hyperparameters across model sizes. Experimentally, we demonstrate significant speedups on nanoGPT training using our algorithm, Scion, without any reliance on Adam. The proposed method is memory-efficient, requiring only one set of model weights and one set of gradients, which can be stored in half-precision. The code is available at https://github.com/LIONS-EPFL/scion .
hub tools
citation-role summary
background 5
citation-polarity summary
roles
background 5polarities
background 5representative citing papers
Gradient clipping eliminates the dependence on maximum delay in the oracle complexity of asynchronous SGD, yielding convergence in expectation and high probability under sub-Weibull gradient noise.
Muon outperforms Adam by reducing curvature penalty via lower Normalized Directional Sharpness, as shown via Taylor approximation on LLM training and proven on stylized quadratic problems with heterogeneous curvature.
A new step size rule lets boosted stochastic Frank-Wolfe match ordinary stochastic Frank-Wolfe rates on nonconvex and quasar-convex problems and deliver faster empirical convergence on sparse logistic regression and quantum tomography.
Regularized Muon induces a damped Hamiltonian flow on probability measures over matrix parameters, yielding exponential convergence under gradient dominance assumptions.
Spectral clipping of leading singular values in gradient matrices stabilizes SGD for non-convex problems with heavy-tailed noise and achieves the optimal convergence rate O(K^{(2-2α)/(3α-2)}).
Muon succeeds by guaranteeing local step-size optimality rather than by tracking any ideal global geometry, as random-spectrum and quasi-norm variants match its performance on language models.
Intrinsic Muon provides closed-form linear maximization oracles on multiple Riemannian matrix manifolds for unitarily invariant norms, with convergence rates depending only on manifold dimension or rank.
A unified stochastic convergence theory is developed for adaptive preconditioned first-order methods including AdaGrad variants, Shampoo, and Muon in nonconvex optimization.
Muon-MVR2 attains the optimal anytime convergence rate of ~O(T^{-1/3}) in stochastic non-convex settings under horizon-free schedules.
Convergence rates are derived for Muon-type methods with inexact LMO in the degenerate case under novel assumptions and layer-wise (L^0, L^1)-smoothness for non-convex and star-convex objectives with weight decay.
OptMuon combines orthogonalized momentum with trajectory-dependent AdaGrad-Norm adaptation to obtain expected-stationarity rates of order T^{-1/2} + sigma^{1/2}T^{-1/4} or T^{-1/2} + sigma^{1/3}T^{-1/3} that reduce to near-optimal deterministic first-order rates in the zero-noise regime.
Introduces a class of asynchronous adaptive first-order methods and establishes O(1/sqrt t) convergence (up to logs) for non-convex stochastic optimization under reasonable assumptions.
HORST uses non-commutative operator composition and a hyperbolic mirror map to combine stability from adaptive optimizers with L1 sparsity bias, outperforming AdamW across sparsity levels on vision and language tasks.
Introduces Distance-Adaptive Muon, Scale-Calibrated Muon, and Distance-Free Muon with stationarity and O(1/T) objective-gap guarantees, shown to match or improve fixed-scale Muon on GPT-124M and ViT-Tiny models.
Rescaled ASGD recovers convergence to the true global objective by rescaling worker stepsizes proportional to computation times, matching the known time lower bound in the leading term under non-convex smoothness and bounded heterogeneity.
Full finetuning with the pretraining optimizer reduces forgetting compared to other optimizers or LoRA while achieving comparable new-task performance.
Manifold constraints via the new MACRO optimizer independently bound activation scales and enforce rotational equilibrium in LLM pre-training, subsuming RMS normalization and decoupled weight decay while delivering competitive performance with convergence guarantees.
SUDA-Muon modularizes decentralized Muon via the SUDA template, proving a topology-separated convergence rate of O((1+σ/√N)K^{-1/4}) in nuclear-norm geometry while establishing that tracking-before-polarization is required to avoid non-stationary fixed points and that local-polarize-then-average is
MuonEq introduces pre-orthogonalization equilibration schemes that improve Muon optimizer performance during large language model pretraining.
Preconditioned matrix norms unify steepest descent, quasi-Newton, and adaptive optimizers, revealing SGD, Adam, Muon, KL-Shampoo, SOAP, and SPlus as special cases and enabling new methods MuAdam and MuAdam-SANIA that are competitive in experiments.
SCALE matches Adam performance in LLM pretraining from 60M to 7B parameters by combining column-wise gradient normalization with last-layer-only momentum, using 35-45% of Adam's memory.
Muon's convergence rate depends on an average Hessian curvature along its update directions, which can be much smaller than the worst-case Lipschitz constant when Hessians are low-rank.
AdaNAGED combines zeroth-order gradient-free training, automatic parameter adaptation, and LMO-based non-Euclidean geometry with claimed convergence guarantees, demonstrated on OPT-1.3B fine-tuning.
citing papers explorer
-
Dead-Direction Conditioners: Gauge-Equivariant Preconditioning for Deep Networks
Dead-Direction Conditioners provide gauge-equivariant preconditioning by conditioning optimizer state on symmetry orbits, yielding improved resistance to over-training collapse and higher detection of dead directions compared to AdamW and Muon.
-
Clipping Makes Distributed and Federated Asynchronous SGD Robust to Stragglers
Gradient clipping eliminates the dependence on maximum delay in the oracle complexity of asynchronous SGD, yielding convergence in expectation and high probability under sub-Weibull gradient noise.
-
Why Muon Outperforms Adam: A Curvature Perspective
Muon outperforms Adam by reducing curvature penalty via lower Normalized Directional Sharpness, as shown via Taylor approximation on LLM training and proven on stylized quadratic problems with heterogeneous curvature.
-
Boosted Stochastic Frank-Wolfe for Constrained Nonconvex Optimization
A new step size rule lets boosted stochastic Frank-Wolfe match ordinary stochastic Frank-Wolfe rates on nonconvex and quasar-convex problems and deliver faster empirical convergence on sparse logistic regression and quantum tomography.
-
Move on Muon : A Hamiltonian probability gradient flow perspective of Muon optimizer
Regularized Muon induces a damped Hamiltonian flow on probability measures over matrix parameters, yielding exponential convergence under gradient dominance assumptions.
-
Gradient Clipping Beyond Vector Norms: A Spectral Approach for Matrix-Valued Parameters
Spectral clipping of leading singular values in gradient matrices stabilizes SGD for non-convex problems with heavy-tailed noise and achieves the optimal convergence rate O(K^{(2-2α)/(3α-2)}).
-
Muon is Not That Special: Random or Inverted Spectra Work Just as Well
Muon succeeds by guaranteeing local step-size optimality rather than by tracking any ideal global geometry, as random-spectrum and quasi-norm variants match its performance on language models.
-
Intrinsic Muon: Spectral Optimization on Riemannian Matrix Manifolds
Intrinsic Muon provides closed-form linear maximization oracles on multiple Riemannian matrix manifolds for unitarily invariant norms, with convergence rates depending only on manifold dimension or rank.
-
A unified convergence theory for adaptive first-order methods in the nonconvex case, including AdaNorm, full and diagonal AdaGrad, Shampoo and Muo
A unified stochastic convergence theory is developed for adaptive preconditioned first-order methods including AdaGrad variants, Shampoo, and Muon in nonconvex optimization.
-
On the Convergence of Muon and Beyond
Muon-MVR2 attains the optimal anytime convergence rate of ~O(T^{-1/3}) in stochastic non-convex settings under horizon-free schedules.
-
Convergence Analysis of Muon-type Methods with Inexact LMO in the Degenerate Case
Convergence rates are derived for Muon-type methods with inexact LMO in the degenerate case under novel assumptions and layer-wise (L^0, L^1)-smoothness for non-convex and star-convex objectives with weight decay.
-
OptMuon: Closed-Loop Orthogonalized Momentum Methods for Stochastic Optimization with Zero-Noise Optimality
OptMuon combines orthogonalized momentum with trajectory-dependent AdaGrad-Norm adaptation to obtain expected-stationarity rates of order T^{-1/2} + sigma^{1/2}T^{-1/4} or T^{-1/2} + sigma^{1/3}T^{-1/3} that reduce to near-optimal deterministic first-order rates in the zero-noise regime.
-
Stochastic convergence of parallel asynchronous adaptive first-order methods
Introduces a class of asynchronous adaptive first-order methods and establishes O(1/sqrt t) convergence (up to logs) for non-convex stochastic optimization under reasonable assumptions.
-
HORST: Composing Optimizer Geometries for Sparse Transformer Training
HORST uses non-commutative operator composition and a hyperbolic mirror map to combine stability from adaptive optimizers with L1 sparsity bias, outperforming AdamW across sparsity levels on vision and language tasks.
-
Distance-Aware Muon: Adaptive Step Scaling for Normalized Optimization
Introduces Distance-Adaptive Muon, Scale-Calibrated Muon, and Distance-Free Muon with stationarity and O(1/T) objective-gap guarantees, shown to match or improve fixed-scale Muon on GPT-124M and ViT-Tiny models.
-
Rescaled Asynchronous SGD: Optimal Distributed Optimization under Data and System Heterogeneity
Rescaled ASGD recovers convergence to the true global objective by rescaling worker stepsizes proportional to computation times, matching the known time lower bound in the leading term under non-convex smoothness and bounded heterogeneity.
-
Optimizer-Model Consistency: Full Finetuning with the Same Optimizer as Pretraining Forgets Less
Full finetuning with the pretraining optimizer reduces forgetting compared to other optimizers or LoRA while achieving comparable new-task performance.
-
Demystifying Manifold Constraints in LLM Pre-training
Manifold constraints via the new MACRO optimizer independently bound activation scales and enforce rotational equilibrium in LLM pre-training, subsuming RMS normalization and decoupled weight decay while delivering competitive performance with convergence guarantees.
-
SUDA-Muon: Structural Design Principles and Boundaries for Fully Decentralized Muon
SUDA-Muon modularizes decentralized Muon via the SUDA template, proving a topology-separated convergence rate of O((1+σ/√N)K^{-1/4}) in nuclear-norm geometry while establishing that tracking-before-polarization is required to avoid non-stationary fixed points and that local-polarize-then-average is
-
MuonEq: Balancing Before Orthogonalization with Lightweight Equilibration
MuonEq introduces pre-orthogonalization equilibration schemes that improve Muon optimizer performance during large language model pretraining.
-
Preconditioned Norms: A Unified Framework for Steepest Descent, Quasi-Newton and Adaptive Methods
Preconditioned matrix norms unify steepest descent, quasi-Newton, and adaptive optimizers, revealing SGD, Adam, Muon, KL-Shampoo, SOAP, and SPlus as special cases and enabling new methods MuAdam and MuAdam-SANIA that are competitive in experiments.
-
Memory-Efficient LLM Pretraining via Minimalist Optimizer Design
SCALE matches Adam performance in LLM pretraining from 60M to 7B parameters by combining column-wise gradient normalization with last-layer-only momentum, using 35-45% of Adam's memory.
-
On the Convergence Analysis of Muon
Muon's convergence rate depends on an average Hessian curvature along its update directions, which can be much smaller than the worst-case Lipschitz constant when Hessians are low-rank.
-
Zero-order Parameter-free Optimization for LMO-based Methods: Novel Approach for Efficient Fine-tuning
AdaNAGED combines zeroth-order gradient-free training, automatic parameter adaptation, and LMO-based non-Euclidean geometry with claimed convergence guarantees, demonstrated on OPT-1.3B fine-tuning.
-
Muon Learns More Robust and Transferable Features than Adam
Muon learns more robust and transferable features than Adam and SGD, shown via corruption robustness tests, transfer experiments, layer-wise probes, effective rank measurements, and a theoretical proof on margins in a multi-component classification problem.
-
Convergence of Spectral Descent for Non-smooth Optimization
Proves linear convergence of Spectral Descent (SD) and Truncated SD for non-smooth convex problems under stated conditions, sublinear rates for regularized versions via Frank-Wolfe, and recovery guarantees for robust low-rank matrix recovery.
-
Anytime Training with Schedule-Free Spectral Optimization
SF-NorMuon is a new schedule-free spectral optimizer that closes the gap with tuned AdamW on 125M-772M parameter models across 1-8x Chinchilla horizons while providing stationarity guarantees.
-
MiMuon: Mixed Muon Optimizer with Improved Generalization for Large Models
MiMuon is a hybrid optimizer that achieves a generalization error bound of O(1/N) independent of the small singular-value gap that limits the original Muon bound, while retaining the same O(1/T^{1/4}) convergence rate.
-
Pion: A Spectrum-Preserving Optimizer via Orthogonal Equivalence Transformation
Pion is an optimizer that preserves the singular values of weight matrices in LLM training by applying orthogonal equivalence transformations.
-
AdaMeZO: Adam-style Zeroth-Order Optimizer for LLM Fine-tuning Without Maintaining the Moments
AdaMeZO adapts Adam moment estimates to zeroth-order LLM fine-tuning without extra memory storage, outperforming MeZO with up to 70% fewer forward passes.
-
Communication-Efficient Gluon in Federated Learning
Compressed Gluon variants using unbiased/contraction compressors and SARAH-style variance reduction achieve convergence guarantees and lower communication costs in federated learning under layer-wise smoothness.
-
A Muon-Accelerated Algorithm for Low Separation Rank Tensor Generalized Linear Models
LSRTR-M integrates Muon updates into the LSRTR algorithm for tensor GLMs, achieving faster convergence, lower estimation errors on synthetic linear/logistic/Poisson models, and competitive performance with better efficiency on Vessel MNIST 3D classification.
-
HTMuon: Improving Muon via Heavy-Tailed Spectral Correction
HTMuon modifies Muon to produce heavier-tailed updates and weight spectra via HT-SR theory, yielding up to 0.98 lower perplexity on LLaMA pretraining and serving as a plug-in for other Muon variants.
-
Low-rank Orthogonalization for Large-scale Matrix Optimization with Applications to Foundation Model Training
Proposes low-rank orthogonalization and derives low-rank Muon and MSGD variants that outperform standard Muon on GPT-2 and LLaMA pretraining while providing iteration complexity bounds.
-
A Note on Stability for Orthogonalized Matrix Momentum with Client Sampling
Derives finite-round upper-tail guarantee on population-empirical gap for client-sampled orthogonalized matrix momentum under heterogeneous data, with Lipschitz condition on the orthogonalizer.
-
Harnessing non-adversarial robustness in large language models
Debiasing via fine-tuning can enhance LLM robustness to semantically neutral prompt perturbations by addressing perturbation-induced bias in neural network outputs.