AI Paper Atlas

Research map / Optimization / Optimizers

Adaptive Batch Sizes Using Non-Euclidean Gradient Noise Scales for Stochastic Sign and Spectral Descent

Hiroki Naganuma, Shagun Gupta, Youssef Briki, Ioannis Mitliagkas, Irina Rish, Parameswaran Raman, Hao-Jun Michael Shi

ICML 2026 · Optimization › Optimizers · approach: SGD momentum and step-size schedules · 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
use non-Euclidean (l1) GNS instead of Euclidean GNS for adaptive batch (ablation)signSGD · C4 160M Llama 3validation loss: 3.8906 -> 3.8396Improved
use non-Euclidean GNS instead of Euclidean GNS (ablation)specSGD · C4 160M Llama 3validation loss: 3.7436 -> 3.7285Improved
use non-Euclidean GNS instead of Euclidean GNS (ablation)Muon · C4 160M Llama 3validation loss: 3.3181 -> 3.3061Improved
adaptive batch size vs constant B=64signSGD · C4 160M Llama 3validation loss: 4.1390 -> 3.8396Improved
adaptive batch size vs constant B=256specSGD · C4 160M Llama 3validation loss: 3.9404 -> 3.7285Improved
adaptive batch size starting at B=64Muon · C4 160M Llama 3steps to baseline loss: 66.77% fewer stepsImproved
adaptive batch size starting at B=64Signum · C4 160M Llama 3steps to baseline loss: 66.61% fewer stepsImproved
adaptive batch size starting at B=64signSGD · C4 1B Llama 3steps to baseline loss: 31.84% fewer stepsImproved
adaptive batch size starting at B=64AdamW · C4 1B Llama 3validation loss: 2.7701 -> 2.8023Hurt
adaptive batch size starting at B=64Signum · C4 1B Llama 3steps to baseline loss: 12.11% fewer stepsImproved
adaptive batch size with matched-norm GNSSGD · ImageWoofvalidation loss: 1.4829 -> 1.3179Improved
adaptive batch size with matched-norm GNSNormalizedSGD · ImageWoofvalidation loss: 1.1979 -> 1.3286Hurt

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