Smooth Interpolation for Improved Discrete Graph Generative Models

Authors: Yuxuan Song, Juntong Shi, Jingjing Gong, Minkai Xu, Stefano Ermon, Hao Zhou, Wei-Ying Ma

ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct extensive experiments and ablation studies on several benchmarks including both abstract and 2D molecules. The empirical results show that the Graph BFN can consistently achieve superior or competitive performance with significantly higher training and sampling efficiency.
Researcher Affiliation Academia 1Institute for AI Industry Research (AIR), Tsinghua University 2Department of Computer Science and Technology, Tsinghua University 3University of Southern California, USA 4Stanford University, USA. Correspondence to: Hao Zhou <EMAIL>.
Pseudocode Yes Algorithm 1 Discrete Variable Bayesian Flow Algorithm 2 Sampling procedure with Adaptive Flowback Algorithm 3 Training procedure for one step
Open Source Code Yes The code is available at https: //github.com/Algo Mole/Graph BFN.
Open Datasets Yes We conduct extensive experiments on both the abstract graph datasets and 2D molecule datasets to compare the performance of Graph BFN against several competitive graph generation baselines... The datasets, dataset splits, and evaluation metrics of QM9 with explicit hydrogen and MOSES follow from Vignac et al. (2022), those of SBM, Planar, QM9 with implicit hydrogen, and ZINC250k follow from Martinkus et al. (2022), and those of Community-small, Ego-small, and Protein follow from Jo et al. (2022).
Dataset Splits Yes The datasets, dataset splits, and evaluation metrics of QM9 with explicit hydrogen and MOSES follow from Vignac et al. (2022), those of SBM, Planar, QM9 with implicit hydrogen, and ZINC250k follow from Martinkus et al. (2022), and those of Community-small, Ego-small, and Protein follow from Jo et al. (2022).
Hardware Specification Yes For all of our experiments, we use NVIDIA Ge Force RTX 3090 with 24GB memory to train and evaluate our models.
Software Dependencies No The models are trained with the Adam W optimizer (Loshchilov & Hutter, 2019) until convergence.
Experiment Setup Yes The most sensitive hyperparameters of Graph BFN are βE and βV that control the speed of information transmission in the generation process. We search them in the discrete linear space {1.0, 2.0, , 10.0}.