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..
Asymptotic theory of SGD with a general learning-rate
Authors: Or Goldreich, Ziyang Wei, SOHAM BONNERJEE, Jiaqi Li, Wei Biao Wu
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | All theoretical findings are corroborated by extensive simulations across diverse settings. ... Our theoretical results are substantiated by extensive numerical exercises. Section 3.2 focuses on linearly decaying schedules... Section 3.3 examines cosine schedules... Some additional numerical exercises can be found in Appendix C. |
| Researcher Affiliation | Academia | Or Goldreich Department of Statistics University of Chicago Chicago, IL 60637 EMAIL Ziyang Wei Department of Statistics University of Chicago Chicago, IL 60637 EMAIL Soham Bonnerjee Department of Statistics University of Chicago Chicago, IL 60637 EMAIL Jiaqi Li Department of Statistics University of Chicago Chicago, IL 60637 EMAIL Wei Biao Wu Department of Statistics University of Chicago Chicago, IL 60637 EMAIL |
| Pseudocode | No | The paper provides mathematical formulations of the SGD algorithm (Equation 1.1) and cyclostationary process, but it does not include any structured pseudocode or algorithm blocks with step-by-step instructions typical of pseudocode. |
| Open Source Code | Yes | All the code files are available in Git Hub. |
| Open Datasets | Yes | To demonstrate the validity of our empirical evaluation beyond elementary linear regression cases, we conducted additional experiments on the MNIST dataset using a high-dimensional classification task. |
| Dataset Splits | No | The paper mentions using the MNIST dataset in Section C.5 but does not explicitly describe how the dataset was split into training, validation, or test sets. It only mentions 'trained a multiclass logistic regression model via stochastic gradient descent (SGD)'. |
| Hardware Specification | No | The paper's 'NeurIPS Paper Checklist' states under Question 8: 'The experiments are lightweight and run quickly on a modern laptop.' No specific hardware details (like GPU/CPU models, memory) are provided within the main text or appendices for the experiments. |
| Software Dependencies | No | The paper does not explicitly mention any specific software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions, or other libraries). |
| Experiment Setup | Yes | All the experiments are based on the following simple linear regression model: yi = θ(0) + θ(1)xi + εi, εi N(0, 1) i.i.d., θ = (θ(0), θ(1)) R2. The true parameter vector is fixed at θ = (2, 3) throughout all experiments. For all the subsequent simulation studies, we initialize the SGD chain at (0, 0). ... We test three fixed values, η = 0.1, 0.05, and 0.01, and track their performance over 104 SGD iterations. ... we simulate SGD with learning rates of the form ηt = η0t β using two values of β: 0.505 and 0.75, each tested with base rates η0 = 0.1, 0.05, 0.01. ... We perform online SGD for n = 104 iterations with η0 {0.01, 0.05, 0.1} and T = 3, averaging results over niter = 500 independent trials. |