Research map / Architectures and efficiency / Attention and transformers
Muon Outperforms Adam in Tail-End Associative Memory Learning
Shuche Wang, Fengzhuo Zhang, Jiaxiang Li, Cunxiao Du, Chao Du, Tianyu Pang, Zhuoran Yang, Mingyi Hong, Vincent Y. F. Tan
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 Muon instead of Adam on all weights | 160M NanoGPT, non-gated FFN · Validation loss (10000 steps, non-gated FFN) | validation loss: 3.9242 -> 3.5654 | Improved |
| add Muon on VO weights (Adam elsewhere) vs Muon on QK only (ablation) | 160M NanoGPT, non-gated FFN · Validation loss (10000 steps, non-gated FFN) | validation loss: 3.8925 -> 3.7644 | Improved |
| add Muon on FFN to VO (VO+FFN vs VO only) (ablation) | 160M NanoGPT, non-gated FFN · Validation loss (10000 steps, non-gated FFN) | validation loss: 3.7644 -> 3.5858 | Improved |
| add Muon on QK to VO+FFN (full Muon vs VO+FFN) (ablation) | 160M NanoGPT, non-gated FFN · Validation loss (10000 steps, non-gated FFN) | validation loss: 3.5858 -> 3.5654 | Improved |
| add Muon on W_o to V+FFN (ablation) | 160M NanoGPT, non-gated FFN · Validation loss (10000 steps, non-gated FFN) | validation loss: 3.6702 -> 3.5858 | Improved |
| add Muon on W_v to O+FFN (ablation) | 160M NanoGPT, non-gated FFN · Validation loss (10000 steps, non-gated FFN) | validation loss: 3.6042 -> 3.5858 | Improved |
| use Muon on W_out instead of W_in only (ablation) | 160M NanoGPT, non-gated FFN · Validation loss (10000 steps, non-gated FFN) | validation loss: 3.7170 -> 3.7023 | Improved |
| add Muon on W_out to VO (vs W_in) (ablation) | 160M NanoGPT, non-gated FFN · Validation loss (10000 steps, non-gated FFN) | validation loss: 3.6778 -> 3.6054 | Improved |
| add Muon on VO+FFN instead of VO only (ablation) | 160M NanoGPT, gated FFN · Validation loss (10000 steps, gated FFN) | validation loss: 3.6874 -> 3.5312 | Improved |
| add Muon on all weights vs VO+FFN (ablation) | 160M NanoGPT, gated FFN · Validation loss (10000 steps, gated FFN) | validation loss: 3.5312 -> 3.5125 | Improved |
| add Muon on W_gate and W_in and W_out to VO (vs W_in+W_gate) (ablation) | 160M NanoGPT, gated FFN · Validation loss (10000 steps, gated FFN) | validation loss: 3.5681 -> 3.5482 | Improved |
| use Muon on all weights instead of Adam | Heavy-tail knowledge task, group 15 · 10000 steps, non-gated FFN | first token accuracy: 0.264 -> 0.976 | Improved |
Closest papers by meaning
- Sharp Capacity Scaling of Spectral Optimizers in Learning Associative Memory (NeurIPS 2026 · similarity 0.91)
- Towards Understanding the Power and Limits of the Muon Optimizer: A River-Valley Perspective (NeurIPS 2026 · similarity 0.90)
- Heavy-Tailed Class Imbalance and Why Adam Outperforms Gradient Descent on Language Models (NeurIPS 2024 · similarity 0.88)
- AMUSE: Anytime Muon with Stable Gradient Evaluation (NeurIPS 2026 · similarity 0.87)
- Adam-mini: Use Fewer Learning Rates To Gain More (ICLR 2025 · similarity 0.87)
- On the Recall Scaling Laws in Mamba: A Theoretical and Mechanistic Study via Hashing (NeurIPS 2026 · similarity 0.86)
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.
