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..
Least squares variational inference
Authors: Yvann Le Fay, Nicolas Chopin, Simon Barthelmé
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Finally, we numerically demonstrate LSVI s performance on various tasks, including logistic regression, discrete variable selection, and Bayesian synthetic likelihood, showing results competitive with state-of-the-art methods, even when gradients are unavailable. |
| Researcher Affiliation | Academia | Yvann Le Fay1, Nicolas Chopin1, , Simon Barthelmé2 1 ENSAE, CREST, IP Paris 2 GIPSA-Lab, CNRS EMAIL EMAIL |
| Pseudocode | Yes | Algorithm 1 Generic LSVI (any family Q) |
| Open Source Code | Yes | We provide a Python package supporting GPU parallelisation via JAX to replicate the experiments: https://github.com/ylefay/LSVI. |
| Open Datasets | Yes | The Sonar (CC BY 4.0 License) and the Census Income (CC BY 4.0 License) datasets are available in the UCI repository while the Pima dataset (CC0: Public Domain License) is in the example datasets of Python package particles (License MIT v0.4, [53, Ch. 1]) and MNIST (CC BYSA 3.0 License) is available at https://github.com/pjreddie/mnist-csv-png. |
| Dataset Splits | No | For all datasets except MNIST, the prior π(β) is a zero-mean Gaussian distribution with diagonal covariance matrix, and the covariances are set to 25 for all the other covariates, except for the intercept, for which it is set to 400. For the MNIST dataset, the prior is a Gaussian distribution with zero-mean and covariance matrix 25In. We use the following standard [e.g., 43] pre-processing strategy for Pima, Sonar and Census-Income datasets: we add an intercept, and we rescale the covariates so that non-binary predictors are centred with standard deviation 0.5, and the binary predictors are centred 0 and range 1. For the third dataset (MNIST dataset), we restrict ourselves to the binary classification problem by selecting pictures labelled 0 or 8. The gray-scale features which range between 0 and 255 are normalised to be between 0 and 1. No intercept is added. For the Census Income dataset, the categorical variables are mapped using one-hot encoding. While preprocessing is detailed, explicit dataset splits for training/testing are not provided. |
| Hardware Specification | Yes | The hardware specifications are CPU AMD EPYC 7702 64-Core Processor and GPU NVIDIA A100-PCIE-40GB, except for SONAR, Census and MNIST datasets where EPYC 7713 and NVIDIA A100-PCIE-80GB were used. |
| Software Dependencies | Yes | All the experiments were conducted using Python 3.13, jax 0.5 with GPU support, Cuda 12.5, and using float64. |
| Experiment Setup | Yes | Initialisations, schedules and number of samples The initialisation distributions for all datasets except MNIST are standard normal distributions. The initialisation for the MNIST dataset is N(0, e 2In). The learning schedules (εt) are obtained via Algorithm 4 with specific inputs (u2, εt) summarised in Table 3 along with the number of samples N. |