PulseAugur
EN
LIVE 15:08:06

New optimizers DMuon and HiMuon boost AI training efficiency · 6 sources tracked

By PulseAugur Editorial · [7 sources] ·

Researchers have developed two new optimization techniques, DMuon and Hierarchical Muon (HiMuon), to improve the efficiency of matrix-orthogonalization-based optimizers like Muon. DMuon integrates into existing training pipelines, offering significant speedups in step time for foundation and large language models, bringing latency close to AdamW levels. HiMuon, on the other hand, uses a tiled approach to Newton-Schulz updates, reducing computational work and enabling efficient GPU utilization for transformer training. Additionally, Tensorion is introduced as a tensor-aware generalization of Muon, extending its capabilities to higher-order tensors and showing promise in computer vision tasks. AI

IMPACT These advancements in optimization techniques could lead to faster and more efficient training of large-scale AI models, particularly in areas like foundation models and computer vision.

RANK_REASON Multiple research papers introducing new optimization techniques for deep learning.

Read on arXiv cs.LG →

AI-generated summary · Google Gemini · from 7 sources. How we write summaries →

New optimizers DMuon and HiMuon boost AI training efficiency · 6 sources tracked

COVERAGE [7]

  1. arXiv cs.LG TIER_1 English(EN) · Vincent Chen, Starrick Liu, Regis Cheng, Dance Yang, Shalfun Li, Ryan Yu, Lucy Liang, Hang Su, Roy Gan, Hao Wang, Qian Wang ·

    DMuon: Efficient Distributed Muon Training with Near-Adam Overhead

    arXiv:2606.27153v1 Announce Type: cross Abstract: Matrix-orthogonalization-based optimizers, exemplified by Muon, have demonstrated strong convergence behavior across a wide range of modern deep learning workloads. The matrix-aware updates offer a compelling alternative to conven…

  2. arXiv cs.LG TIER_1 English(EN) · Ziyuan Tang, Tianshi Xu, Yousef Saad, Yuanzhe Xi ·

    Hierarchical Muon: Tiled Newton-Schulz Updates for Efficient Muon Optimization

    arXiv:2606.27216v1 Announce Type: cross Abstract: Muon-type optimizers construct update directions for dense neural-network weights by applying a finite Newton-Schulz map to momentum-gradient matrices. For an $H \times W$ matrix, with $r=\min\{H,W\}$ and $s=\max\{H,W\}$, $K$ step…

  3. arXiv cs.LG TIER_1 English(EN) · Yuanzhe Xi ·

    Hierarchical Muon: Tiled Newton-Schulz Updates for Efficient Muon Optimization

    Muon-type optimizers construct update directions for dense neural-network weights by applying a finite Newton-Schulz map to momentum-gradient matrices. For an $H \times W$ matrix, with $r=\min\{H,W\}$ and $s=\max\{H,W\}$, $K$ steps of the full-matrix Newton-Schulz update require …

  4. arXiv cs.LG TIER_1 English(EN) · Qian Wang ·

    DMuon: Efficient Distributed Muon Training with Near-Adam Overhead

    Matrix-orthogonalization-based optimizers, exemplified by Muon, have demonstrated strong convergence behavior across a wide range of modern deep learning workloads. The matrix-aware updates offer a compelling alternative to conventional element-wise optimization, particularly as …

  5. arXiv cs.LG TIER_1 English(EN) · Vladimir Bogachev, Vladimir Aletov, Alexander Molozhavenko, Sergei Kudriashov, Maxim Rakhuba ·

    Tensorion: A Tensor-Aware Generalization of the Muon Optimizer

    arXiv:2606.25975v1 Announce Type: new Abstract: Common first-order optimizers, such as Adam, implicitly treat each parameter block as an unstructured vector, which disregards the multilinear weight structure present in many modern machine learning models. Recent work has shown th…

  6. arXiv cs.LG TIER_1 English(EN) · Maxim Rakhuba ·

    Tensorion: A Tensor-Aware Generalization of the Muon Optimizer

    Common first-order optimizers, such as Adam, implicitly treat each parameter block as an unstructured vector, which disregards the multilinear weight structure present in many modern machine learning models. Recent work has shown that exploiting matrix structure can improve optim…

  7. Hugging Face Daily Papers TIER_1 English(EN) ·

    Tensorion: A Tensor-Aware Generalization of the Muon Optimizer

    Common first-order optimizers, such as Adam, implicitly treat each parameter block as an unstructured vector, which disregards the multilinear weight structure present in many modern machine learning models. Recent work has shown that exploiting matrix structure can improve optim…

RELATED ENTITIES

RELATED TOPICS