Research map / Architectures and efficiency / RNNs and normalization
Spectra: Rethinking Optimizers for LLMs Under Spectral Anisotropy
Zhendong Huang, Hengjie Cao, Fang Dong, Ruijun Huang, Mengyi Chen, Yifeng Yang, Xin Zhang, Anrui Chen, Mingzhi Dong, Yujiang Wang, Jinlong Hou, Qin Lv and 5 more
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 |
|---|---|---|---|
| replace AdamW with Spectra | Qwen3-0.6B · Downstream average | average accuracy: 51.55 -> 52.96 | Improved |
| replace Muon with Spectra | Qwen3-0.6B · Downstream average | average accuracy: 52.07 -> 52.96 | Improved |
| replace AdamW with Spectra | LLaMA3-8B · Downstream average | average accuracy: 56.39 -> 58.01 | Improved |
| replace Muon with Spectra | LLaMA3-8B · Downstream average | average accuracy: 57.35 -> 58.01 | Improved |
| replace AdamW with Spectra | Qwen3-0.6B · DCLM validation loss | loss: 2.83 -> 2.77 | Improved |
| replace Muon with Spectra | LLaMA3-8B · DCLM validation loss | loss: 2.60 -> 2.59 | Improved |
| replace AdamW with Spectra | Qwen3-0.6B · Computational efficiency | optimizer state memory: -49.25% | Improved |
| replace AdamW with Spectra | Qwen3-0.6B · Computational efficiency | step time (ms): 5297 -> 5259 | Improved |
| replace Muon Newton-Schulz with cached power iteration | Muon · Latency (4096, 14336) | latency (ms): 9.1465 -> 1.7799 | Improved |
| increase power iterations from 1 to 8 | Spectra power iteration · Latency (4096, 14336) | latency (ms): 1.7799 -> 7.2676 | Hurt |
| apply frequency-normalized loss (FreqNorm) to gradients (ablation) | Qwen3-0.6B gradient spectrum · Gradient singular spectrum | largest singular value: >25% drop | Improved |
| shuffle tokens within sentences (ablation) | Qwen3-0.6B gradient spectrum · Gradient singular spectrum | leading singular values: up to 75% increase | Hurt |
Closest papers by meaning
- Sharp Capacity Scaling of Spectral Optimizers in Learning Associative Memory (NeurIPS 2026 · similarity 0.89)
- Towards Understanding the Power and Limits of the Muon Optimizer: A River-Valley Perspective (NeurIPS 2026 · similarity 0.88)
- Dynamic Sparse No Training: Training-Free Fine-tuning for Sparse LLMs (ICLR 2024 · similarity 0.87)
- Pion: A Spectrum-Preserving Optimizer via Orthogonal Equivalence Transformation (NeurIPS 2026 · similarity 0.87)
- SUMO: Subspace-Aware Moment-Orthogonalization for Accelerating Memory-Efficient LLM Training (NeurIPS 2025 · similarity 0.87)
- R-Sparse: Rank-Aware Activation Sparsity for Efficient LLM Inference (ICLR 2025 · 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.
