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..

Equivariance Everywhere All At Once: A Recipe for Graph Foundation Models

Authors: Ben Finkelshtein, Ismail Ilkan Ceylan, Michael Bronstein, Ron Levie

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We validate our approach through extensive experiments on 29 real-world node classification datasets, demonstrating both strong zero-shot empirical performance and consistent improvement as the number of training graphs increases.
Researcher Affiliation Collaboration Ben Finkelshtein University of Oxford Ismail Ilkan Ceylan TU Wien / AITHYRA/ University of Oxford Michael Bronstein University of Oxford / AITHYRA Ron Levie Technion Israel Institute of Technology
Pseudocode No The paper does not contain any explicit pseudocode blocks or algorithms. It describes methods using mathematical formulations and textual descriptions rather than structured algorithmic steps.
Open Source Code Yes Experiments are executed on a single NVIDIA L40 GPU, and publicly accessible at: https://github.com/benfinkelshtein/Equivariance Everywhere.
Open Datasets Yes We use an ensemble of 29 node classification datasets and their respective official splits. Specifically, we use roman-empire, amazon-ratings, minesweeper, questions and tolokers from [30], cora, citeseer and pubmed from [49], chameleon and squirrel from [36], cornell, wisconsin, texas and actor from [29], full-dblp and full-cora from [3], wiki-attr and blogcatalog from [48], wiki-cs from [27], co-cs, co-physics, computers and photo from [39], brazil, usa and europe from [34], last-fm-asia and deezer from [35], and arxiv from [17].
Dataset Splits Yes We use an ensemble of 29 node classification datasets and their respective official splits. Table 8: Statistics of the 28 node node classification datasets. ... Train/Val/Test Ratios (%) We train TS-GNNs by randomly masking the labels of 50% of the labeled nodes at each epoch and optimizing the model to predict the masked labels from the node features and the remaining visible labels.
Hardware Specification Yes Experiments are executed on a single NVIDIA L40 GPU, and publicly accessible at: https://github.com/benfinkelshtein/Equivariance Everywhere.
Software Dependencies No The paper mentions graph learning libraries such as Py G [10] and DGL [44] (Section H) but does not provide specific version numbers for these or other software components used in the experiments.
Experiment Setup Yes We optimize all models using the Adam optimizer, with detailed hyperparameter settings provided in Section J. Table 10: Hyperparameters used in Table 1.