Research map / Optimization / Optimizers
From Muon to Gluon: Bridging Theory and Practice of LMO-based Optimizers for LLMs
Artem Riabinin, Egor Shulgin, Kaja Gruntkowska, Peter Richtárik
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use larger step size for the embedding layer than for transformer blocks (non-uniform layer-wise scaling) in unScion (ablation) | unScion · NanoGPT-124M on FineWeb | final validation loss: best at rho2 = 50 and rho3 = 3000, i.e., t_i = 0.018 and t_p = 1.08 | Improved |
| fit layer-wise (L0,L1) smoothness instead of constant-Li per layer (ablation) | smoothness model of NanoGPT · NanoGPT-124M trajectory smoothness | relative fit error: substantially smaller relative error than constant-Li fits | Improved |
| measure smoothness along AdamW instead of Scion trajectories (ablation) | layer-wise (L0,L1) smoothness model · NanoGPT-124M | fit of trajectory smoothness: fitted L1 of the embedding layer is approximately 20-30 times smaller than in other groups | No clear effect |
| use specialized (non-Euclidean) norms instead of the Euclidean norm (ablation) | smoothness fit · NanoGPT-124M | trajectory smoothness approximation | Improved |
Closest papers by meaning
- From SGD to Muon: Adaptive Optimization via Schatten-p Norms (NeurIPS 2026 · similarity 0.93)
- Towards Understanding the Power and Limits of the Muon Optimizer: A River-Valley Perspective (NeurIPS 2026 · similarity 0.89)
- Training Deep Learning Models with Norm-Constrained LMOs (ICML 2025 · similarity 0.89)
- Tensorion: A Tensor-Aware Generalization of the Muon Optimizer (NeurIPS 2026 · similarity 0.88)
- MuLoCo: Muon is a Practical Inner Optimizer for DiLoCo (ICML 2026 · similarity 0.88)
- AMUSE: Anytime Muon with Stable Gradient Evaluation (NeurIPS 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
