AI Paper Atlas

Research map / Optimization / Implicit bias and generalization

How Muon’s Spectral Design Benefits Generalization: A Study on Imbalanced Data

Bhavya Vasudeva, Puneesh Deora, Yize Zhao, Vatsal Sharan, Christos Thrampoulidis

ICLR 2026 · Optimization › Implicit bias and generalization · approach: Explicit and implicit regularization interplay

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
use spectral norm update (Muon) instead of Frobenius (NMD) or max-norm (Signum) updatesMLP · Colored-MNIST (99% digit-colour correlation), minority groupstest accuracyImproved
use Muon instead of NMD and SignumResNet-18/50 · CIFAR-10/100 STEP imbalance 20:1, minority classestest accuracyImproved
use SpecGD instead of GD (early stopped)linear model, squared loss · Gaussian mixture, class imbalanceminority-class and balanced loss gap: minority gap >= mu t/4, balanced gap >= mu t/2Improved
use SpecGD instead of normalized GDlinear model, squared loss · Gaussian mixture, class imbalanceminority-class and balanced loss gap: gaps >= mu t/2Improved
use SpecGD instead of NGD/SignGD (early stopped)linear model, cross-entropy · heavy-tailed imbalance, 20 classesclass-balanced and worst-class accuracyImproved
increase depth of SpecGD-trained linear modelSpecGD · deep linear modelgap between component saturation timesImproved
use Muon instead of SGDResNet-34 · Dominoes MNIST-CIFAR, worst-group accuracyworst-group and decoded worst-group accuracyImproved
use Muon instead of AdamResNet-34 · Dominoes MNIST-CIFAR, worst-group accuracyworst-group accuracyNo clear effect
use Muon/Shampoo instead of SGDBERT-base · MultiNLI fine-tuningworst-group accuracyImproved
use Muon instead of AdamBERT-base · MultiNLI fine-tuningworst-group accuracyHurt
use Muon instead of AdamResNet-50 · CelebA fine-tuningworst-group accuracyNo clear effect
use Muon instead of SGD and Adamsmall Transformer · TinyStories rare and frequent tokenstop-1 and top-5 accuracyImproved

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