Research map / Optimization / Nonconvex and smooth optimization
LiMuon: Light and Fast Muon Optimizer for Large Models
Feihu Huang, Yuning Luo, Songcan Chen
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 |
|---|---|---|---|
| increase LiMuon rank from 4 to 16 (ablation) | LiMuon · Mamba-130M WikiText-103 | validation perplexity: 79.12 -> 53.84 | Improved |
| use full rank instead of rank 16 (ablation) | LiMuon · Mamba-130M WikiText-103 | validation perplexity: 53.84 -> 47.78 | Improved |
| increase LiMuon rank from 4 to full rank (ablation) | LiMuon · Mamba-130M | memory (GB): 18.86 -> 22.80 | Hurt |
| increase LiMuon rank from 4 to 8 (ablation) | LiMuon · Qwen2.5-0.5B MiniPile | validation perplexity: 92.06 -> 54.87 | Improved |
| use full rank instead of rank 16 (ablation) | LiMuon · Qwen2.5-0.5B MiniPile | validation perplexity: 46.77 -> 40.83 | Improved |
| use LiMuon rank=8 instead of Muon | optimizer · Qwen2.5-0.5B | memory (GB): 54.14 -> 52.55 | Improved |
| use LiMuon full rank instead of Muon | optimizer · Qwen2.5-0.5B | validation perplexity: 67.60 -> 40.83 | Improved |
| use LiMuon full rank instead of Muon | optimizer · ViT Tiny-ImageNet | validation top-1 accuracy (%): 47.87 -> 48.04 | Improved |
| use LiMuon rank 4 instead of Muon | optimizer · ViT Tiny-ImageNet | training top-1 accuracy (%): 78.96 -> 64.73 | Hurt |
| use LiMuon full rank instead of Muon | optimizer · ViT | time to target: 7min 6s -> 7min 1s | Improved |
| use low-rank LiMuon (rank 4) instead of Muon | optimizer · ViT | time to target: 7min 6s -> 12min 4s | Hurt |
| use LiMuon rank 4 instead of Muon | optimizer · Mamba-130M | validation perplexity: 71.27 -> 79.12 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- SUMO — SUMO: Subspace-Aware Moment-Orthogonalization for Accelerating Memory-Efficient LLM Training (NeurIPS 2025) · better on 4 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- On the Convergence Analysis of Muon (NeurIPS 2026 · similarity 0.88)
- Powering Up Zeroth-Order Training via Subspace Gradient Orthogonalization (NeurIPS 2026 · similarity 0.87)
- AMUSE: Anytime Muon with Stable Gradient Evaluation (NeurIPS 2026 · similarity 0.87)
- Sharp Capacity Scaling of Spectral Optimizers in Learning Associative Memory (NeurIPS 2026 · similarity 0.87)
- Muon Does Not Converge on Convex Lipschitz Functions (NeurIPS 2026 · similarity 0.87)
- Towards Understanding the Power and Limits of the Muon Optimizer: A River-Valley Perspective (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.
