Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..

Stab-SGD: Noise-Adaptivity in Smooth Optimization with Stability Ratios

Authors: David A. R. Robin, Killian Bakong, Kevin Scaman

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We validate these statements with experiments in convex and deep learning scenarios in Sec. 4.
Researcher Affiliation Academia David A. R. Robin INRIA ENS Paris PSL Research UniversityKillian Bakong INRIA ENS Paris PSL Research UniversityKevin Scaman INRIA ENS Paris PSL Research University
Pseudocode Yes Algorithm 1: Inline Stab-SGD
Open Source Code Yes The source code to reproduce all experiments of this section and the next is available online at https://www.github.com/robindar/2025-NeurIPS_Stab-SGD.
Open Datasets Yes We perform experiments on the CIFAR-10 image classification dataset [Krizhevsky, 2009]
Dataset Splits Yes We use batches of size 128 sampled without replacement for each epoch (391 batches / epoch). We restrict the hyperparameter search for log10(η0) to a grid from 3 to +1 by steps of 0.5, informed by choices in the original reference. We use an ℓ2 2 weight decay with λ = 10 4 for all runs.
Hardware Specification No Average runtime of 4h to 5h per seed on GPU.
Software Dependencies No The paper does not specify any software dependencies with version numbers.
Experiment Setup Yes We compare with the aforementioned original ResNet publication, which uses a learning rate 10 1 for 32k iterations, then 10 2 for the next 16k and 10 3 for the last 16k, totaling 64k iterations (thresholds depicted by dashed vertical lines). We use batches of size 128 sampled without replacement for each epoch (391 batches / epoch). We restrict the hyperparameter search for log10(η0) to a grid from 3 to +1 by steps of 0.5, informed by choices in the original reference. We use an ℓ2 2 weight decay with λ = 10 4 for all runs. We run Algorithm 1 with η0 = 10+1, with the configuration κ = 10 1, γ = 1 and ζ = 100.