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..

Decreasing Entropic Regularization Averaged Gradient for Semi-Discrete Optimal Transport

Authors: Ferdinand Genans, Antoine Godichon-Baggioni, François-Xavier Vialard, Olivier Wintenberger

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our theoretical findings are supported by numerical experiments that validate the effectiveness of DRAG and highlight its practical advantages. 5 Numerical experiments
Researcher Affiliation Academia Sorbonne Université, CNRS, LPSM1 Université Gustave Eiffel, CNRS, LIGM2 Wolfgang Pauli Institute3 EMAIL EMAIL
Pseudocode Yes Algorithm 1 DRAG Parameters: (γ1, a, b, C) Initialize g0 C, g0 = g0, ε0 = 1. for k = 1 to t do Xk µ gk =Proj C gk 1 γk ghεk 1(Xk, gk 1) gk = 1 k+1gk + k k+1gk 1 εk = k a end for return gt
Open Source Code Yes All the code to reproduce our experiments is provided in Python Notebooks and attached in a zip file in the supplementary materials.
Open Datasets Yes We replicate their pipeline on toy datasets from their repository [27], replacing their ADAMbased solver with DRAG, using the same number of samples. As we can see, while both solvers yield good results on the "swissroll" target data, DRAG outperforms the ADAM solver on the "spiralarms" data, being able to almost completely generate it, whereas ADAM shows poorer coverage.
Dataset Splits No The paper describes using synthetic datasets like uniform distributions and mentions using 't iid samples from the source measure' in an online setting. It does not provide explicit training/test/validation splits in the conventional sense, as it focuses on an online algorithm for optimal transport estimation rather than a supervised learning task.
Hardware Specification No No specific computing resources are required for our experiments; they can be run on any modern machine, even without a GPU. The only exception is the Neural OT algorithm, which does require a GPU.
Software Dependencies No The paper mentions providing code in "Python Notebooks" in the supplementary materials, but it does not specify any version numbers for Python, or any specific libraries (e.g., PyTorch, TensorFlow, NumPy) and their corresponding versions.
Experiment Setup Yes In all figures and experiments, we set the parameters of DRAG to εt = 0.1/ta, a = 0.33, b = 2/3, γ1 = Diam(Supp(µ)). Our numerical investigation found that our parameter selection is robust without further hypertuning. For this comparison, we fixed the parameters of DRAG to (M, 1/3, 2/3) and ran the algorithm for t = 105 iterations. The parameters for Adam were set to β1 = 0.9, β2 = 0.999, and λ = 10 3 (learning rate/weight decay).