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..
Momentum Multi-Marginal Schrödinger Bridge Matching
Authors: Panagiotis Theodoropoulos, Augustinos Saravanos, Evangelos A. Theodorou, Guan-Horng Liu
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We empirically validate the computational and performance benefits of our method. [...] We evaluate the performance of 3MSBM on synthetic and real-world trajectory inference tasks, such as Lotka-Volterra (Sec. 4.1), ocean current in the Gulf of Mexico (Sec. 4.2), single-cell sequencing (Sec. 4.4), and the Beijing air quality data (Sec. 4.3). We compare against state-of-the-art methods explicitly designed to incorporate multi-marginal settings, [...] using the same metric in all datasets: the Sliced Wasserstein Distance (SWD). Additional results for these tasks with more metrics and baselines are provided in Appendix E. |
| Researcher Affiliation | Collaboration | Panagiotis Theodoropoulos1, Augustinos D. Saravanos1, Evangelos A. Theodorou1, , Guan-Horng Liu2, 1Georgia Institute of Technology, 2FAIR at Meta, Equal advising |
| Pseudocode | Yes | Algorithm 1 Momentum Multi-Marginal Schrödinger Bridge Matching (3MSBM) 1: Input: Marginals q(x0), q(x1), . . . , q(x N), R, σ, K, T 2: Initialize aθ t, q({xn}) := q(x0) q(x1) q(x N), and v0 N(0, I) 3: repeat 4: for j = 0 to J do 5: Calculate at|{ xn} using Eq. (6) for t from 0 to T 6: v N sdeint(x0, v0, at|{xn}], σ, K, T) 7: Calculate at|{ xn} using Eq. (6) for t from T to 0 8: v0 sdeint(x N, v N, at|{ xn}, σ, K, T) 9: end for 10: Update aθ t, from Eq. (9) using at|{ xn} 11: Sample new x0, x1, . . . , x N from aθ t 12: until converges |
| Open Source Code | No | Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [No] Justification: We have not released the code yet, but we plan to do so very soon. |
| Open Datasets | Yes | We first consider a synthetic dataset generated by the Lotka Volterra (LV) equations [Goel et al., 1971], which model predator-prey interactions through coupled nonlinear dynamics. We used the dataset from [Shen et al., 2024] with the 5 training and 4 validation time points, with 50 observations per time point. [...] The dataset contains ocean-current snapshots of the velocity field around a vortex in the Gulf of Mexico (Go M). Similarly to the LV dataset, we used the big vortex dataset provided in [Shen et al., 2024] [...]. Lastly, we demonstrate the efficacy of 3MSBM to infer trajectories in high-dimensional spaces. In particular, we use the Embryoid Body (EB) stem cell differentiation tracking dataset, [...] We used the preprocessed dataset provided by [Tong et al., 2020, Moon et al., 2019] [...]. To further study the capacity of 3MSBM to effectively infer missing values, we also tested it in the Beijing multi-site air quality data set [Chen, 2017]. |
| Dataset Splits | Yes | The generated dataset consists of 9 marginals in total; the even-numbered indices are used to train the model (i.e., t0, t2, t4, t6, t8), and the remainder of the time points are used to assess the efficacy of our model to impute and infer the missing time points. [...] Out of the total 9 marginals, the even-indexed time points (i.e., t0, t2, t4, t6, t8) are used for training, and the remaining are left out to evaluate the model s ability to impute and infer missing temporal states. [...] For the imputation task, we omitted the data at t2, t5, t8, and t11, while the remaining snapshots formed the training set. [...] The training set consists of the even-indexed time-steps (i.e., t0, t2, t4), while the rest are used as the validation set. |
| Hardware Specification | Yes | Experiments were run on an RTX 4090 GPU with 24 GB of VRAM. |
| Software Dependencies | No | For all experiments with our 3MSBM, we employed the Res Net architectures from Chen et al. [2023b], Dockhorn et al. [2021]. We used the Adam W optimizer and applied Exponential Moving Averaging with a decay rate of 0.999. All results are averaged over 5 random seeds, with means and standard deviations reported in Section 4 and the tables below. |
| Experiment Setup | Yes | The hyperparameter selection for the LV with our method were: the diffusion coefficient was set to σ = 0.3, and the learning rate was 10 4. [...] The hyperparameters used for the Go M experiment with our method were: a batch size of 32 for the matching, the diffusion coefficient was set to σ = 0.3, and the learning rate was set equal to 2 10 4. [...] The hyperparameters used for the Beijing air quality experiment with our method were: a total number of samples of 1000 were used, with a batch size of 64 for the matching, the diffusion coefficient was set to σ = 0.2, and the learning rate was set to 5 10 5. [...] The hyperparameters used for every EB experiment with our method were: a total number of samples of 1000 were used, with a batch size of 64 for the matching, the diffusion coefficient was set to σ = 0.1, and the learning rate was set to 10 4. |