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..
Conditional Gradient Methods with Standard LMO for Stochastic Simple Bilevel Optimization
Authors: Khanh-Hung (Bruce) Giang-Tran, Soroosh Shafiee, Nam Ho-Nguyen
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results on over-parametrized regression and dictionary learning tasks demonstrate the practical advantages of our approach over existing methods, confirming our theoretical findings. |
| Researcher Affiliation | Academia | Khanh-Hung Giang-Tran Cornell University EMAIL Soroosh Shafiee Cornell University EMAIL Nam Ho-Nguyen The University of Sydney EMAIL |
| Pseudocode | Yes | Algorithm 1: Iteratively Regularized Stochastic Conditional Gradient (IR-SCG) Algorithm. ... Algorithm 2: Iteratively Regularized Finite-Sum Conditional Gradient (IR-FSCG) Algorithm. |
| Open Source Code | Yes | To ensure reproducibility, all source codes are made available at https://github.com/brucegiang/CG-StoBilvl. |
| Open Datasets | Yes | We use the same training and validation datasets, (Atr, btr) and (Aval, bval), from the Wikipedia Math Essential dataset [43], as in [4]. |
| Dataset Splits | Yes | We use the same training and validation datasets, (Atr, btr) and (Aval, bval), from the Wikipedia Math Essential dataset [43], as in [4]. |
| Hardware Specification | No | This paper only studies relatively small-scale problems compared to other machine learning areas, and each individual experiment can run on an ordinary laptop. |
| Software Dependencies | No | For SBCGI and SBCGF, a linear optimization oracle over the ℓ1-norm ball intersecting with a half-space is required, which we employ CVXPY [11] to solve the problems, similar to the implementation in [4, Appendix F.1]. |
| Experiment Setup | Yes | For IR-SCG, presented in Algorithm 1, we set σt = ς(t + 1) 1/4, where ς = 10, along with αt = 2/(t + 2). For IR-FSCG, presented in Algorithm 2, we set S = q = n , σt = ς(max{t, q} + 1) 1/2, where ς = 10, along with αt = 2/(t + 1) for every t q and αt = log(q)/q for every t < q. |