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..
PDPO: Parametric Density Path Optimization
Authors: Sebastian Gutierrez Hernandez, Peng Chen, Hao-Min Zhou
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirically, we find that using 3 5 control points of the spline interpolation suffices to accurately resolve both multimodal and high-dimensional problems. We demonstrate that PDPO can flexibly accommodate a wide range of potential terms, including those modeling obstacles, mean-field interactions, stochastic control, and higher-order dynamics. Our method outperforms existing state-of-the-art approaches in benchmark tasks, demonstrating superior computational efficiency and solution quality. Source code https://github.com/Sebas Gut Hdz/PDPO/tree/main . 4 Experiments We conduct a comprehensive evaluation of our method across a variety of benchmark scenarios to demonstrate its accuracy, efficiency, flexibility, and scalability. In all experiments, we compare against GSBM [20], the current state-of-the-art, which has been shown to outperform earlier approaches. |
| Researcher Affiliation | Academia | Sebastian Gutierrez Hernandez1 Peng Chen2 Haomin Zhou1 1School of Mathematics, Georgia Institute of Technology 2School of Computational Science and Engineering, Georgia Institute of Technology EMAIL |
| Pseudocode | Yes | Algorithm 1 Path optimization Algorithm 2 Coupling optimization Algorithm 3 PDPO Algorithm 4 Parameterized MLP Forward Pass |
| Open Source Code | Yes | Source code https://github.com/Sebas Gut Hdz/PDPO/tree/main . 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: [Yes] Justification: We provide a copy of our code in the supplemental materials. |
| Open Datasets | Yes | SCC (S-Curve with Congestion) from [17]: Particles navigate around two obstacles arranged in an "S"-shaped configuration while interacting with one another. VNEFI (V-Neck with Entropy and Fisher Information) from [20]: This stochastic optimal control (SOC) problem involves particles navigating a narrow channel while minimizing entropy, see Figure 11b. Gaussian Mixture obstacle (GMM) from [20]: In this benchmark, particles must move between multimodal source and target distributions while avoiding a Gaussian mixture-shaped obstacle, see Figure 13 for the density path in the appendix. The definition and source code were taken from [20]. |
| Dataset Splits | No | For estimating the action, we use Monte Carlo integration with 3,000 samples and compute the time integral using the trapezoidal rule with a step size of t = 1/50. Each experiment is repeated across three random trials with different seeds. The paper describes the use of Monte Carlo samples for estimation and mentions the trapezoidal rule for time integration, but it does not specify conventional dataset splits (e.g., training/validation/test sets). The problems are defined by source and target distributions, not by splitting a large dataset. |
| Hardware Specification | Yes | We implement our method in Py Torch [28], and run all experiments on an AMD 7543 CPU + NVIDIA RTX A6000 GPU. |
| Software Dependencies | No | We implement our method in Py Torch [28], and run all experiments on an AMD 7543 CPU + NVIDIA RTX A6000 GPU. The Wasserstein-2 distance is approximated using the POT library [12] with 3,000 samples. The paper mentions PyTorch and the POT library but does not provide specific version numbers for these software components. |
| Experiment Setup | Yes | Additional implementation and experimental details are provided in Appendix F. In Table 4 see the boundary conditions ρ0 and ρ1, and Table 4 the hyper-parameters for the PDPO algorithm. Table 4 explicitly lists hyperparameters such as K, N, M, Architecture, Epochs, Coupling opt steps, Path opt. steps, Geodesic warmup steps, α, and (κ0, κ1, κ2) for each experiment. |