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..
Differentiation Through Black-Box Quadratic Programming Solvers
Authors: Connor Magoon, Fengyu Yang, Noam Aigerman, Shahar Kovalsky
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Comprehensive benchmark experiments demonstrate d QP s robustness and scalability, particularly highlighting its advantages in large-scale sparse problems. |
| Researcher Affiliation | Academia | 1 Department of Mathematics, University of North Carolina at Chapel Hill, 2 Université de Montréal and Mila |
| Pseudocode | Yes | Algorithm 1 d QP: Differentiation through Black-box Quadratic Programming Solvers |
| Open Source Code | Yes | Our open-source code is available at https://github.com/cwmagoon/d QP. |
| Open Datasets | Yes | We tested d QP on the QP Benchmark suite [31], focusing on 129 large-scale sparse problems from the standard Maros-Meszaros (MM) dataset [79], which are widely used as stress tests for QP solvers. |
| Dataset Splits | No | The paper evaluates performance on benchmark problems and randomly generated QPs, organizing them by dimension and number of problems per dimension. However, it does not specify explicit training/test/validation dataset splits typically used for machine learning model training, as the work focuses on evaluating a solver framework rather than training a predictive model on these datasets. |
| Hardware Specification | Yes | All experiments in this section were run on a Macbook Air with Apple M2 chips, 8 cores, and 16GB RAM. The geometry experiments were run on an Intel(R) Core(TM) i7-8850H CPU @ 2.60GHz with 6 cores. |
| Software Dependencies | Yes | We implement d QP (Algorithm 1) as a fully differentiable Py Torch module [88], providing an intuitive interface for integrating differentiable QPs into machine learning and bi-level optimization workflows. To ensure modularity, the forward pass supports any QP solver interfaced via the open-source qpsolvers library [32] |
| Experiment Setup | Yes | For the forward, we set the absolute residual tolerance to ϵabs = 10 6 and the active constraint tolerance to ϵJ = 10 5. We run each problem separately with batch size 1. Experimental results are averaged over 5 independent samples. For the upper-level optimization, we use the Adam optimizer with learning rate 10 2 [67]. |