AI Paper Atlas

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

ICML 2026 · Architectures and efficiency › RNNs and normalization · approach: Neural scaling laws · 0 citations (OpenAlex)

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 changedWhereMeasuredEffect
replace AdamW with SpectraQwen3-0.6B · Downstream averageaverage accuracy: 51.55 -> 52.96Improved
replace Muon with SpectraQwen3-0.6B · Downstream averageaverage accuracy: 52.07 -> 52.96Improved
replace AdamW with SpectraLLaMA3-8B · Downstream averageaverage accuracy: 56.39 -> 58.01Improved
replace Muon with SpectraLLaMA3-8B · Downstream averageaverage accuracy: 57.35 -> 58.01Improved
replace AdamW with SpectraQwen3-0.6B · DCLM validation lossloss: 2.83 -> 2.77Improved
replace Muon with SpectraLLaMA3-8B · DCLM validation lossloss: 2.60 -> 2.59Improved
replace AdamW with SpectraQwen3-0.6B · Computational efficiencyoptimizer state memory: -49.25%Improved
replace AdamW with SpectraQwen3-0.6B · Computational efficiencystep time (ms): 5297 -> 5259Improved
replace Muon Newton-Schulz with cached power iterationMuon · Latency (4096, 14336)latency (ms): 9.1465 -> 1.7799Improved
increase power iterations from 1 to 8Spectra power iteration · Latency (4096, 14336)latency (ms): 1.7799 -> 7.2676Hurt
apply frequency-normalized loss (FreqNorm) to gradients (ablation)Qwen3-0.6B gradient spectrum · Gradient singular spectrumlargest singular value: >25% dropImproved
shuffle tokens within sentences (ablation)Qwen3-0.6B gradient spectrum · Gradient singular spectrumleading singular values: up to 75% increaseHurt

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy