Are Expressive Encoders Necessary for Discrete Graph Generation?

A research team led by Jay Revolinsky, Harry Shomer, and Jiliang Tang has published a paper challenging conventional wisdom in graph generation. Their work introduces GenGNN, a modular message-passing framework that questions whether highly expressive neural backbones like transformers are truly necessary for discrete graph generation tasks. The framework achieves remarkable results, with diffusion models using GenGNN reaching over 90% validity on Tree and Planar datasets while operating at 2-5x faster inference speeds compared to traditional approaches.

For practical applications like molecule generation, the system demonstrates exceptional performance. When combined with the DiGress diffusion model, GenGNN achieves 99.49% validity in generated molecules. The researchers conducted systematic ablation studies revealing that residual connections are crucial for mitigating oversmoothing in complex graph structures. Through scaling analyses, they applied a principled metric-space approach to investigate learned diffusion representations, providing new insights into whether GNNs can serve as expressive neural backbones for discrete diffusion models.

The findings suggest that simpler, more efficient graph neural network architectures can match or exceed the performance of complex transformer models for specific graph generation tasks. This has significant implications for computational chemistry, drug discovery, and materials science where generating valid molecular structures quickly and accurately is essential. The research opens new avenues for optimizing graph generation pipelines while reducing computational overhead.