Research map / Optimization / Optimizers
An Exploration of Non-Euclidean Gradient Descent: Muon and its Many Variants
Michael Crawshaw, Chirag Modi, Mingrui Liu, Robert M. Gower
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 stale nuclear norm approximation (ablation) | MuonMax · FineWeb 1B tokens | final validation loss: 3.791 -> 3.768 | Improved |
| use stale nuclear norm approximation (ablation) | MuonAdam-Momo · FineWeb 1B tokens | final validation loss: 3.551 -> 3.554 | No clear effect |
| use stale nuclear norm approximation (ablation) | Scion-Momo · FineWeb 1B tokens | final validation loss: 3.592 -> 3.590 | No clear effect |
| use stale nuclear norm approximation (ablation) | MuonMax-Momo · FineWeb 1B tokens | final validation loss: 3.576 -> 3.580 | No clear effect |
| use stale nuclear norm approximation (ablation) | MuonMax-Momo · FineWeb 1B tokens | extra wall-clock time vs MuonAdam: 11% -> 5% | Improved |
| add Momo model-based step size | MuonAdam · FineWeb 1B tokens | final validation loss: 3.604 -> 3.551 | Improved |
| add Momo model-based step size | MuonMax · FineWeb 1B tokens | final validation loss: 3.791 -> 3.576 | Improved |
| add Momo model-based step size | MuonAdam · FineWeb 1B tokens, LR sweep | proportion of LRs with loss < 3.65 (%): 25% -> 62.5% | Improved |
| use MuonMax-Momo instead of Muon-Adam or Scion | MuonMax-Momo · SlimPajama 1B tokens (GPT2-Large) | LR robustness: 3.13 to 3.24 | Improved |
| use MuonMax-Momo instead of MuonAdam | MuonMax-Momo · SlimPajama 6B tokens (GPT2-Large) | final validation loss | Improved |
| lower the Momo loss lower bound F* from 3.2 to 0.0 (ablation) | MuonMax-Momo · FineWeb 1B tokens | final validation loss: 3.58 -> 3.61 | Hurt |
Closest papers by meaning
- Towards Understanding the Power and Limits of the Muon Optimizer: A River-Valley Perspective (NeurIPS 2026 · similarity 0.91)
- On the Convergence Analysis of Muon (NeurIPS 2026 · similarity 0.90)
- Tensorion: A Tensor-Aware Generalization of the Muon Optimizer (NeurIPS 2026 · similarity 0.90)
- AMUSE: Anytime Muon with Stable Gradient Evaluation (NeurIPS 2026 · similarity 0.90)
- From SGD to Muon: Adaptive Optimization via Schatten-p Norms (NeurIPS 2026 · similarity 0.89)
- MuLoCo: Muon is a Practical Inner Optimizer for DiLoCo (ICML 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.
