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..
HollowFlow: Efficient Sample Likelihood Evaluation using Hollow Message Passing
Authors: Johann Flemming Gloy, Simon Olsson
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We validate Hollow Flow by training BGs on two different systems of increasing size. For both systems, the sampling and likelihood evaluation time decreases dramatically, following our theoretical scaling laws. For the larger system we obtain a 102 speed-up, clearly illustrating the potential of Hollow Flow-based approaches for high-dimensional scientific problems previously hindered by computational bottlenecks. Experiments on multi-particle systems validate Hollow Flow s scalability, bridging a critical gap in high-dimensional generative modelling for the sciences. |
| Researcher Affiliation | Academia | 1Department of Computer Science and Engineering, Chalmers University of Technology and University of Gothenburg SE-41296 Gothenburg, Sweden. EMAIL |
| Pseudocode | Yes | Algorithm 1 Hollow message passing |
| Open Source Code | Yes | The code and the models are available here: https://github.com/olsson-group/hollowflow. |
| Open Datasets | Yes | We used the same training data as in [23] available at https://osf.io/srqg7/?view_only= 28deeba0845546fb96d1b2f355db0da5. |
| Dataset Splits | Yes | For all systems we used 105 randomly selected samples for training as in [23]. Validation was done using 104 of the remaining samples. |
| Hardware Specification | Yes | The training and inference for LJ13 and Alanine Dipeptide was conducted on a NVIDIA Tesla V100 SXM2 with 32GB RAM. The training for LJ55 was conducted on a NVIDIA Tesla A100 HGX with 40GB RAM, inference was performed on a NVIDIA Tesla A40 with 48GB RAM. |
| Software Dependencies | No | All models and training is implemented using Py Torch [71] with additional use of the following libraries: bgflow [1, 24], torchdyn [72], Torch CFM [20, 22] and Sch Net Pack [73, 74]. |
| Experiment Setup | Yes | The training details for LJ13, LJ55 and Alanine Dipeptide are reported in tables 8 to 10. For LJ13, we selected the last model for all k NN experiments and the fully connected baseline experiment, while we selected the model with the lowest validation loss for all remaining experiments for inference. For LJ55 and Alanine Dipeptide we always selected the model with the lowest validation loss for inference. All training was done using the Adam optimizer [76]. The integration of the ODE (eq. (3)) was performed using a fourth order Runge Kutta solver with a fixed step size. We used 20 integration steps. All neural networks in the Pai NN architecture use the Si LU activation function [75]. |