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..
Follow the Energy, Find the Path: Riemannian Metrics from Energy-Based Models
Authors: Louis Bethune, David Vigouroux, Yilun Du, Rufin VanRullen, Thomas Serre, Victor Boutin
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our approach on increasingly complex datasets: synthetic datasets with known data density, rotated character images with interpretable geometry, and high-resolution natural images embedded in a pretrained VAE latent space. Our results show that EBM-derived metrics consistently outperform established baselines, especially in high-dimensional settings. |
| Researcher Affiliation | Academia | Louis Bethune Apple David Vigouroux IRT Saint Exupéry, ANITI, IMT Atlantique Yilun Du Harvard University Rufin Van Rullen CNRS Thomas Serre Brown University Victor Boutin CNRS |
| Pseudocode | Yes | Algorithm 1: Training geodesic interpolant |
| Open Source Code | Yes | The code to reproduce all our experiments is available at https://github.com/Victor Boutin/Riemann EBM. |
| Open Datasets | Yes | We evaluate our approach on increasingly complex datasets: synthetic datasets with known data density, rotated character images with interpretable geometry, and high-resolution natural images embedded in a pretrained VAE latent space. We now evaluate our method on the Animal Faces High Quality (AFHQ) dataset [59] |
| Dataset Splits | No | The paper describes the datasets used and how evaluation trajectories were sampled (e.g., 'averaged over 1,000 geodesics with distinct endpoints' or '50,000 trajectories'), but it does not specify explicit training/validation/test splits for the underlying data used to train the Energy-Based Models. |
| Hardware Specification | Yes | All experiments were conducted on NVIDIA RTX 3090 GPUs (32 GB memory). |
| Software Dependencies | No | The paper mentions using the Adam optimizer [94] but does not specify version numbers for programming languages, libraries (e.g., PyTorch, TensorFlow), or other software dependencies. |
| Experiment Setup | Yes | In all experiments, we use L = 100 Langevin steps with step size α = 1 and noise scale σ = 10^-2. The energy function is optimized using the Adam optimizer [94] with a learning rate of η = 10^-4. ... We trained the model using the Adam optimizer [94] with a learning rate of 1e-4 and a batch size of 128. |