arXiv:2606. 25971v1 Announce Type: new Abstract: Modern neural network training relies on optimizers such as Adam and Muon which act on each weight matrix as a single object.
The paper introduces the Drift Contract, a spectral update geometry for local learning that improves depth robustness and hyperparameter stability. By applying momentum orthogonalization with spectral step scaling to per‑layer updates, the authors achieve consistent performance across a wide range of widths and depths on CIFAR‑10 MLPs, outperforming local Adam and providing a per‑layer, input‑conditioned drift bound. The study also shows that the spectral geometry itself, rather than step‑size rules, drives the observed depth robustness, while a negative result indicates that the stability benefit is limited to non‑normalized layers.
arXiv:2606. 04058v1 Announce Type: cross Abstract: Orthonormalized update rules have rapidly become a leading choice of optimizer for training large language models, with recent open-source state-of-the-art models adopting Muon.
The paper investigates how much spectral detail is necessary for the Muon optimizer, which traditionally uses a flat spectral profile. By analyzing singular modes, the authors find that most modes lie below a noise edge yet align positively with the gradient, leading them to propose BulkBoost—a two‑band spectral reweighting framework that adjusts bulk and spike gains while preserving matrix norms. Experiments across various pre‑training settings show that this low‑dimensional reweighting matches or surpasses fine‑grained spectral profiles, achieving notable loss reductions.
arXiv:2607. 23711v1 Announce Type: new Abstract: LoRA fine-tuning can create intruder dimensions: new leading singular vectors of the updated weight matrix $W+BA$ that are nearly orthogonal to all pretrained singular vectors and that drive catastrophic forgetting.
The paper introduces a physical response-and-memory model for the Muon optimizer, explaining its semi‑orthogonalized momentum update as the maximally dissipative direction under an output‑side safety budget. It treats the weight matrix as a responsive medium with internal stress, showing that momentum corresponds to accumulated stress whose relaxation occurs over multiple timescales—fast and slow. Based on this, the authors propose the Bi‑Maxwell optimizer, which uses a two‑timescale memory kernel and achieves target loss in fewer steps on a public large‑language‑model benchmark.
arXiv:2609.08381v1 Announce Type: cross Abstract: Equivariant networks are commonly trained with Adam, yet recent work reports that matrix-structured optimizers such as Muon can perform better on the...
arXiv:2606. 29176v1 Announce Type: new Abstract: A deep network's loss is invariant to continuous symmetries of its parameters: the logit shift, the ReLU rescaling, the LayerNorm scale, the per-head attention rotation.
The paper evaluates ten deep learning optimizers across models ranging from 0.1B to 1.2B parameters and varying data-to-model ratios, highlighting that fair comparisons require rigorous hyperparameter tuning and end-of-training evaluations. It finds that claimed speedups over AdamW are lower than advertised, dropping to only 1.1× for the largest models, and that intermediate checkpoint comparisons can mislead due to learning‑rate decay. The study shows that matrix‑based optimizers like Muon and Soap achieve the highest speedups, but these benefits diminish as model size increases.
LayerRoPE introduces a depth‑wise magnitude and angular superposition mechanism that replaces per‑layer normalization gains with a single shared vector and depth‑conditioned scalars, reducing parameters and computational cost. Across 16 pre‑trained large language models and a 100B+ token model ladder, LayerRoPE outperforms traditional Pre‑, Post‑, and Peri‑Norm schemes, achieving comparable or better performance with significantly less compute and improved convergence up to 512 layers. The method demonstrates that widening the residual stream, rather than shrinking it, enhances depth stability and learning‑rate sensitivity, and it transfers effectively to latent and vision transformer architectures.
The paper introduces a new closed‑form scaling law that extends Chinchilla’s original formula to handle data‑constrained regimes. It decomposes loss into undercapacity, undertraining, and overfitting components, saturating between an irreducible loss and an uninformed baseline. The authors validate the model on diverse architectures and domains, achieving state‑of‑the‑art RMSE across multiple LLM scaling‑law grids and enabling cost‑aware training allocations.
A deep network's loss is invariant to continuous symmetries of its parameters: the logit shift, the ReLU rescaling, the LayerNorm scale, the per-head attention rotation. Adam's per-coordinate preconditioner drifts along each symmetry orbit, which pulls the trajectory off the symmetry quotient where the optimization lives and blurs the singular-learning rate the quotient makes readable.