AI Paper Atlas

Research map / Optimization / Optimizers

Improved Stochastic Optimization of LogSumExp

Egor Gladin, Alexey Kroshnin, Jia-Jie Zhu, Pavel Dvurechensky

ICML 2026 · Optimization › Optimizers · approach: Stochastic gradient MCMC

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 KL with Safe KL (rho=1e-3) in KL-DROKL-DRO SGD · California housing, lambda=1, |D|=10objective: 20.0 -> 2.1Improved
replace KL with Safe KL (rho=1e-3) with large batchKL-DRO SGD · California housing, lambda=1/5, |D|=1000objective: 9.1 -> 25.3Hurt
use rho=1e-1 instead of 1e-3 (larger rho) (ablation)Safe KL KL-DRO · California housing, lambda=1, |D|=1000objective: 2.5 -> 21.8Hurt
use rho=1e-5 instead of 1e-3 (smaller rho) (ablation)Safe KL KL-DRO · California housing, lambda=5, |D|=10objective: 0.76 -> 1.03Hurt
replace KL with Safe KL (rho=1e-1) in UOT-DROUOT-DRO SGD · MNIST 25% noise, gamma=1, lambda=1objective: 0.81 -> 0.37Improved
use rho=1 instead of 1e-1 (ablation)Safe KL UOT-DRO · MNIST 25% noise, gamma=1/5, lambda=1/5objective: 1.57 -> 4.76Hurt
replace KL with Safe KL (rho=1e-1) in UOT-DRO under 85% label noiseUOT-DRO SGD · MNIST 85% noise, gamma=1, lambda=1objective: 1.26 -> 0.83Improved
replace KL with Safe KL under 85% label noiseUOT-DRO SGD · MNIST 85% noisecross-entropy on hardest class: 2.15 -> 1.87Improved
replace KL with Safe KL (rho=1e-1) in KL-DRO under distribution shiftKL-DRO SGD · ACS PUMS CA to MSRMSE on hardest group (x10^3): 75.8 -> 74.7Improved
replace KL with Safe KL (rho=1e-1) in KL-DRO under distribution shiftKL-DRO SGD · ACS PUMS CA to MStime per epoch (s): 92 -> 99Hurt

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