Research map / Architectures and efficiency / RNNs and normalization
NorMuon: Making Muon more efficient and scalable
Zichong Li, Liming Liu, Chen Liang, Weizhu Chen, Tuo Zhao
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 |
|---|---|---|---|
| add neuron-wise normalization after orthogonalization to Muon | Muon · SlimPajama 1.1B | efficiency gain over Adam (%): 10.43 -> 21.74 | Improved |
| add neuron-wise normalization after orthogonalization to Muon | Muon · SlimPajama 5.4B | efficiency gain over Adam (%): 6.08 -> 13.91 | Improved |
| add neuron-wise normalization to Muon | Muon · Modded-NanoGPT FineWeb 124M | iterations to reach same validation loss: Muon requires 6% more iterations | Improved |
| add neuron-wise normalization to Muon | Muon · Modded-NanoGPT FineWeb 350M | iterations to reach same validation loss: Muon requires 15% more iterations | Improved |
| use neuron-wise instead of coordinate-wise (Adam) normalization after Muon (ablation) | NorMuon · 1.1B pretraining | validation loss | Improved |
| use neuron-wise instead of coordinate-wise (Adam) normalization after Muon (ablation) | NorMuon · 5.4B overhead | optimizer state memory (GB): 40.56 -> 21.19 | Improved |
| apply normalization after orthogonalization rather than before (ablation) | NorMuon · 1.1B pretraining | validation loss | Improved |
| apply normalization to all matrices rather than m>n only (ablation) | NorMuon · 1.1B pretraining | validation loss | Improved |
| add neuron-wise adaptivity rather than only plain normalization of Muon update (ablation) | Muon with normalization · Modded-NanoGPT 350M | validation loss | Improved |
| apply NorMuon to all matrices instead of only m>n ones (ablation) | NorMuon · Modded-NanoGPT 350M | validation loss | Improved |
| add orthogonalization distribution across devices (ablation) | NorMuon · 5.4B overhead | optimizer step time (s): 2.29 -> 0.84 | Improved |
| use NorMuon instead of AdamW | AdamW · 5.4B overhead | training step time (s): 28.73 -> 29.57 | Hurt |
Closest papers by meaning
- Towards Understanding the Power and Limits of the Muon Optimizer: A River-Valley Perspective (NeurIPS 2026 · similarity 0.90)
- From SGD to Muon: Adaptive Optimization via Schatten-p Norms (NeurIPS 2026 · similarity 0.88)
- AMUSE: Anytime Muon with Stable Gradient Evaluation (NeurIPS 2026 · similarity 0.88)
- MuLoCo: Muon is a Practical Inner Optimizer for DiLoCo (ICML 2026 · similarity 0.88)
- Sharp Capacity Scaling of Spectral Optimizers in Learning Associative Memory (NeurIPS 2026 · similarity 0.87)
- On the Convergence Analysis of Muon (NeurIPS 2026 · similarity 0.87)
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.
