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..
Learning Theory for Kernel Bilevel Optimization
Authors: Fares El Khoury, Edouard Pauwels, Samuel Vaiter, Michael Arbel
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We numerically illustrate our theoretical findings on a synthetic instrumental variable regression task. ... In Section 5, we illustrate our theoretical findings with experiments on synthetic data for the instrumental variable regression problem. ... The plots in Figure 2 show the generalization behavior as a function of the number of inner samples n. |
| Researcher Affiliation | Academia | Fares El Khoury Université Grenoble Alpes, Inria, CNRS, Grenoble INP, LJK, 38000 Grenoble, France Edouard Pauwels Toulouse School of Economics, Université Toulouse Capitole, 31080 Toulouse, France Samuel Vaiter CNRS & Université Côte d Azur, Laboratoire J. A. Dieudonné, 06108 Nice, France Michael Arbel Université Grenoble Alpes, Inria, CNRS, Grenoble INP, LJK, 38000 Grenoble, France Correspondence to: EMAIL. |
| Pseudocode | No | The paper describes methods and mathematical formulations in prose and equations. There are no explicitly labeled pseudocode blocks or algorithm figures in the main text or appendices. |
| Open Source Code | Yes | Our code is available at https://github.com/fareselkhoury/KBO. |
| Open Datasets | No | We numerically illustrate our theoretical findings on a synthetic instrumental variable regression task. ... We use the Gaussian kernel and follow the experimental setup of Singh et al. [68], generating synthetic data that remain fixed across all runs. ... We generate synthetic data as follows: x Px, t = 2(1 p x + ϵ), y = ω ϕ(t) + ϵ, where ϵ N(0, 0.025), ω U(0, 1)d, and ϕ(t) = (sin(t + 1), . . . , sin(t + d)) . ... All random variables are fixed across runs for reproducibility. |
| Dataset Splits | No | We vary n between 100 and 5,000, setting m = n. ... The logs of the means of the four quantities across 50 runs are displayed. The paper describes generating synthetic data for each run and varying sample sizes (n and m) but does not define explicit training/test/validation splits for a fixed dataset, which is not applicable given the synthetic data generation approach. |
| Hardware Specification | Yes | We use the JAX framework [18] to run our experiments on an NVIDIA RTX 6000 ADA GPU. |
| Software Dependencies | No | We use the JAX framework [18] to run our experiments on an NVIDIA RTX 6000 ADA GPU. The paper mentions the JAX framework but does not provide a specific version number, which is required for a reproducible description of ancillary software. |
| Experiment Setup | Yes | We set p = 3, d = 4, λ = 0.01, and σ = 0.2. ... We optimize the outer loss in ( d KBO) using gradient descent, where the step size is selected using backtracking line search and ω0 is randomly drawn from U(0, 1)d. The stopping criterion is when b F(ωi) 10 5, where i is the iteration index. |