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..
Wasserstein Convergence of Critically Damped Langevin Diffusions
Authors: Stanislas Strasman, Sobihan Surendran, Claire Boyer, Sylvain Le Corff, Vincent Lemaire, Antonio Ocello
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The benefits of this additional parameterization are demonstrated numerically on challenging synthetic datasets. We illustrate the effect of the regularization parameter ε on the generation quality of CLDs on a simple yet challenging toy dataset. Tables 1, 2 and 3 report the sliced-Wasserstein error for different values of the regularization parameter ε... |
| Researcher Affiliation | Collaboration | 1Sorbonne Université and Université Paris Cité, CNRS, LPSM, F-75005 Paris, France 2LOPF, Califrais Machine Learning Lab, Paris, France 3LMO, Université Paris-Saclay, UMR CNRS 8628, Institut Universitaire de France, Orsay, France 4CREST, ENSAE, Institut Polytechnique de Paris, Palaiseau, France |
| Pseudocode | Yes | Algorithms 1 and 2 show the training and sampling procedures for the CLD-based approaches, respectively. |
| Open Source Code | Yes | Our source code is publicly available here1. 1https://github.com/Sobihan Surendran/CLD |
| Open Datasets | Yes | Dataset. We evaluate the generation quality on the Funnel distribution, which is characterized by a strong imbalance in variance across dimensions and was previously used in Thin et al. (2021). To further illustrate our results, we extend the evaluation to two additional challenging toy datasets (Appendix E.5): MG-25 (a 25-mode, 100-dimensional Gaussian mixture) and Diamonds (a 2dimensional Gaussian mixture with a diamond-shaped geometry). |
| Dataset Splits | Yes | The training set consists of 50 000 samples. For evaluation, we generate 50 000 samples using the Euler Maruyama discretization scheme with N = 1000 steps and compare them against a test set of 50 000 samples. |
| Hardware Specification | No | We would also like to thank SCAI (Sorbonne Center for Artificial Intelligence) for providing the computing clusters. |
| Software Dependencies | No | No specific software dependencies with version numbers are mentioned in the paper. |
| Experiment Setup | Yes | All score networks share the same architecture: a fully connected neural network with three hidden layers of width 512 (see Figure 3). Training is performed using the Adam optimizer to minimize the hybrid score matching objective in (18), with a learning rate of 10 4 over 2000 epochs. The training set consists of 50 000 samples. For evaluation, we generate 50 000 samples using the Euler Maruyama discretization scheme with N = 1000 steps and compare them against a test set of 50 000 samples. Both training and generation are independently repeated five times. |