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..
Schrödinger Bridge Matching for Tree-Structured Costs and Entropic Wasserstein Barycentres
Authors: Samuel Howard, Peter Potaptchik, George Deligiannidis
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We first examine the performance of Tree DSBM in a low-dimensional synthetic example, using the experimental setup of Noble et al. (2023). We compute the (1/3)-barycentre of a moon, spiral, and circle dataset with ̑̑ = 0.1 (recall ̑ = ̑̑̑/2). We compare Tree DSBM ran for 6 IMF iterations against Tree DSB ran for 50 IPF iterations (using checkpoints provided by Noble et al. (2023)). In Figure 2, we plot the obtained samples from both methods, and for comparison also display the barycentre obtained using the in-sample method from Cuturi and Doucet (2014) to give a close approximation to the ground-truth. To quantitatively assess performance, we report the Sinkhorn divergence (Genevay et al., 2018) relative to this ground-truth in Table 2. |
| Researcher Affiliation | Academia | Samuel Howard Department of Statistics University of Oxford Peter Potaptchik Department of Statistics University of Oxford George Deligiannidis Department of Statistics University of Oxford Corresponding author: EMAIL |
| Pseudocode | Yes | Algorithm 1: Tree DSBM Algorithm 2: Tree DSBM for Wasserstein barycentre computation |
| Open Source Code | Yes | Our code is available at https: //github.com/samuel-howard/Tree_SB_Matching_Barycentres. |
| Open Datasets | Yes | We compute the (1/3)-barycentre of a moon, spiral, and circle dataset with ̑̑ = 0.1 (recall ̑ = ̑̑̑/2). MNIST 2,4,6 barycentre We also compare performance of Tree DSBM with Tree DSB on a higher dimensional image dataset, computing the ( 1/3)-barycentre between MNIST digits 2, 4, and 6 (Le Cun et al., 2010). We consider the experimental setup and dataset used in Korotin et al. (2021) (the same dataset was also used previously in Li et al. (2020) and Fan et al. (2021)), which uses Poisson and negative-binomial regressions on a bikerental dataset (Fanaee-T, 2013). Korotin et al. (2022) proposed the Ave, celeba! barycenter benchmark, which consists of 3 distributions of transformed Celeb A faces (Liu et al., 2015) |
| Dataset Splits | Yes | The dataset is 8-dimensional and is split into 5 distinct subsets each of size 100,000. |
| Hardware Specification | Yes | All experiments were ran on a single Nvidia Ge Force RTX 2080Ti GPU. The experiments were conducted on Nvidia A100 GPUs on Google Colaboratory. |
| Software Dependencies | Yes | We implement the Tree DSBM procedure using the JAX framework (Bradbury et al., 2018). Bradbury, J., Frostig, R., Hawkins, P., Johnson, M. J., Leary, C., Maclaurin, D., Necula, G., Paszke, A., Plas, J. V., Wanderman-Milne, S., and Zhang, Q. (2018). JAX: composable transformations of Python+NumPy programs. Version 0.3.13. |
| Experiment Setup | Yes | For Tree DSBM, we use ̑̑ = 0.1 and run for 6 IMF iterations. For training the vector fields, we use 10,000 training steps and a batch size of 4096. We use the Adam optimiser (with default parameters 0.9, 0.999) with learning rate 1e-3 and exponential moving average parameter of 0.99. |