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..
Learning non-equilibrium diffusions with Schrödinger bridges: from exactly solvable to simulation-free
Authors: Stephen Zhang, Michael Stumpf
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In application to a range of problems based on synthetic and real single cell data, we demonstrate that MVOU-OTFM achieves higher accuracy compared to competing methods, whilst being significantly faster to train. |
| Researcher Affiliation | Academia | Stephen Y. Zhang University of Melbourne Michael P. H. Stumpf University of Melbourne |
| Pseudocode | Yes | Algorithm 1 mv OU-OTFM: score and flow matching for mv OU-Schrödinger Bridges |
| Open Source Code | Yes | Code is available at https://github.com/zsteve/mv OU_SBP. |
| Open Datasets | Yes | We next apply our framework to a single-cell RNA sequencing (sc RNA-seq) dataset [4] of the cell cycle |
| Dataset Splits | Yes | We run Algorithm 2 for 5 iterations and we compare in Fig. 4(b) the ground truth vector field to the SB vector field, both the fitted mv OU reference (at the final iterate of Alg. 2) and for a Brownian reference (as the first iterate of Alg. 2 or equivalently the output of [52]). As was found by [44, 58], iterated fitting of a global autonomous vector field for the reference process leverages multi-timepoint information and allows reconstruction of dynamics better adapted to the underlying system. |
| Hardware Specification | Yes | For (mv OU, BM)-OTFM and IPFP, all computations were carried out by CPU (8x Intel Xeon Gold 6254). NLSB and SBIRR computations were accelerated using a single NVIDIA L40S GPU. |
| Software Dependencies | No | We do this using the standard Ridge CV method implemented in the scikit-learn package, which automatically selects the regularisation parameter. |
| Experiment Setup | Yes | We use a batch size of 64 and learning rate 10 2 for 2,500 iterations using the Adam W optimiser. |