SymDrift: One-Shot Generative Modeling under Symmetries

SymDrift: One-Shot Generative Modeling under Symmetries – Data and Checkpoints This record contains the processed datasets and pretrained model checkpoints for the paper "SymDrift: One-Shot Generative Modeling under Symmetries" by S. Darouich, V. Tong, L. Pastor-Pérez, T. Bien, L. Mualem and M. Niepert (NeurIPS 2026, arXiv:2605.06140). SymDrift introduces symmetry-aware drifting models for one-shot generation of molecular systems. There are two approaches: symmetrized coordinate-space drifting, and drifting in G-invariant latent embeddings. Inference is up to 40× faster than with multi-step diffusion and flow matching models, with competitive performance. Contents - datasets/: processed versions of GEOM-QM9, GEOM-DRUGS and RDB7, together with the train/validation/test split files. Place the contents under ./data in the code repository so that each dataset lies in ./data/<dataset_name>.- checkpoints/: pretrained SymDrift models for both drifting variants: - coordinate-space models (symmetrized coordinate-space drifting) - embedded-space models (G-invariant embedding drifting) Usage The code, environment setup, and instructions for training, sampling and evaluation are available at: https://github.com/samirdarouich/SymDrift Citation If you use these models, please cite: Darouich, S., Tong, V., Pastor-Pérez, L., Bien, T., Mualem, L., & Niepert, M. (2026). SymDrift: One-Shot Generative Modeling under Symmetries. Advances in Neural Information Processing Systems (NeurIPS). arXiv:2605.06140. The processed datasets in datasets/ are derived from the following works. If you use them, please cite the original sources. GEOM-QM9 and GEOM-DRUGS – conformer data from the GEOM dataset:Axelrod, S., & Gómez-Bombarelli, R. (2022). GEOM, energy-annotated molecular conformations for property prediction and molecular generation. Scientific Data, 9, 185. https://doi.org/10.1038/s41597-022-01288-4 RDB7 – reaction data (reactants, transition states, products):Spiekermann, K., Pattanaik, L., & Green, W. H. (2022). High accuracy barrier heights, enthalpies, and rate coefficients for chemical reactions. Scientific Data, 9, 417. https://doi.org/10.1038/s41597-022-01529-6