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..
Universally Invariant Learning in Equivariant GNNs
Authors: Jiacheng Cen, Anyi Li, Ning Lin, Tingyang Xu, Yu Rong, Deli Zhao, Zihe Wang, Wenbing Huang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirical results demonstrate that our model demonstrates superior completeness and excellent performance with only a few layers, thereby significantly reducing computational overhead while maintaining strong practical efficacy. In this section, we conduct two categories of experiments, totaling seven in number, to validate our theory and methods: a) three toy datasets in 4.1 to assess the expressivity of our models; b) the remaining four in 4.2 evaluate the actual performance of our models. |
| Researcher Affiliation | Collaboration | 1 Gaoling School of Artificial Intelligence, Renmin University of China 4 DAMO Academy, Alibaba Group, Hangzhou, China EMAIL; EMAIL; EMAIL; EMAIL; EMAIL |
| Pseudocode | Yes | For example, we present Algos. 1 and 2, which is based on the four-point positioning principle and operates with complexities of O(N 6) where N is the number of nodes, based on the four-point positioning principle, and further turned it into canonical form for comprehensive embeddings of geometric graphs. Additionally, we theoretically prove that a full-rank basis set of any degree can always be constructed, if the input geometric graph is asymmetric. Building on this foundation, we propose a more efficient algorithm with a complexity of O(N 2) to construct the canonical form specifically for asymmetric graphs. Built upon Algos. 1 and 2, we have derived a canonical form Γ( ) as presented in Algo. 3. Thus, we derive Algo. 4, yielding more efficiency. |
| Open Source Code | Yes | 2Code is available at https://github.com/GLAD-RUC/Uni-EGNN. |
| Open Datasets | Yes | 4.2 Performance on Physical Systems Dataset setup. We conducted four experiments to evaluate our model s performance across different scenarios: ... c) 100-body system [66], d) MD17 dataset [68], and e) Water-3D mini [66; 67], where predictions are made based on initial coordinates and velocities. |
| Dataset Splits | Yes | We utilize 500/2000/2000 samples for training, validation, and testing, respectively. We use 5000/2000/2000 samples for training, validation, and testing like the setting in HEGNN [49]. Due to time constraints, we only evaluate a substantial subset (Water-3D-mini: 3,000/300/300 for train/val/test). |
| Hardware Specification | Yes | We conducted experiments on a single NVIDIA H20 GPU. |
| Software Dependencies | No | The paper mentions software like e3nn [61], EGNN [43], TFN [33], MACE [35], and HEGNN [49] but does not provide specific version numbers for these or other ancillary software components. |
| Experiment Setup | Yes | In both experiments, we configured the batch size to 100, the learning rate to 5 10 4, and the weight decay to 10 12. For the tetrahedron center prediction task, the maximum number of training epochs was set to 300. In the case of the N-body system task, we implemented early stopping with a patience of 100 steps. Each model utilizes at most 4 layers, and the dimension of the hidden embedding is fixed at 64. |