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..
Rao-Blackwell Gradient Estimators for Equivariant Denoising Diffusion
Authors: Vinh Tong, Trung-Dung Hoang, Anji Liu, Guy Van den Broeck, Mathias Niepert
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirically, Orbit Diffusion achieves state-of-the-art results on GEOM-QM9 for molecular conformation generation, improves crystal structure prediction, and advances text-guided crystal generation on the Perov-5 and MP-20 benchmarks. Additionally, it enhances protein designability in protein structure generation. ... Our experiments evaluate the generality and robustness of Orbit Diffusion across diverse generative tasks. We begin with a controlled synthetic setup using a standard diffusion model (Section 4.1) and consider various isometry groups... We then extend our method to Flow Matching (Section 4.2) and to diffusion models with non-standard forward processes (Section 4.3). Finally, we apply Orbit Diffusion to a non-equivariant denoiser, demonstrating its effectiveness without architectural symmetry (Section 4.4). ... Table 1: Synthetic experiment results: RMSD to the closest target in { 1, 1} and W2 distance to ground-truth distribution. ... Table 2: Molecular conformer generation performance on GEOMQM9. ... Table 3: Text-guided CSP with TGDMat. ... Table 4: Crystal Structure Prediction (CSP). ... Table 5: Protein Structure Generation. |
| Researcher Affiliation | Academia | 1University of Stuttgart, 2IMPRS-IS, 3UCLA, 4University of Bern, 5National University of Singapore EMAIL |
| Pseudocode | Yes | Algorithm 1: Orbit Diffusion with RB. ... The pseudo-code for the Rao-Blackwell estimator with SNIS is shown on the right. |
| Open Source Code | Yes | Code is available at https://github.com/vinhsuhi/Orbit-Diffusion.git. |
| Open Datasets | Yes | We evaluate on the GEOM-QM9 dataset (Axelrod & Gomez-Bombarelli, 2022), respecting two key symmetries: invariance under global 3D rotations and equivariance to graph automorphisms... We evaluate our method on two CSP benchmarks: Perov-5 (Castelli et al., 2012a,b) and MP-20 (Jain et al., 2013). |
| Dataset Splits | Yes | We follow the same train/validation/test split as in (Ganea et al., 2021; Jing et al., 2022), consisting of 106,586 / 13,323 / 1,000 molecules, respectively. |
| Hardware Specification | Yes | All models were trained on a single NVIDIA Ge Force RTX 4090 GPU. ... Our fine-tuning setup uses 4 A100 GPUs for 24 hours on the same dataset, whereas the original training employed 96 GPUs. |
| Software Dependencies | No | The paper mentions software components like "pynauty library" and "pymatgen library" without specifying version numbers. It also refers to "Protein MPNN (Dauparas et al., 2022)" and "ESMFold (Lin et al., 2023)" which are models/software, but again, without version numbers for replication. |
| Experiment Setup | Yes | During training, we apply symmetry-aware sampling by uniformly sampling 50 automorphisms and 200 SO(3) rotations per molecule, including the identity. These are applied to both 2D graphs and 3D conformers. All other settings follow ETFLOW; see Appendix C.1 for details. ... All other training settings, including optimizer configurations and learning rate schedules, follow the defaults of ETFLOW. ... TGDMat was trained for 1,500 epochs, while Diff CSP was trained for 500 epochs. ... Our fine-tuning setup uses 4 A100 GPUs for 24 hours on the same dataset, with a batch size of 8 and 32 gradient accumulation steps. ... For each generated backbone, we produce 8 candidate sequences using Protein MPNN (Dauparas et al., 2022) with a sampling temperature of 0.1. |