Title: Enhancing Novel Graph Generation Through Efficient, Structure-Aware Autoregressive Approaches

The creation of realistic and diverse graphs represents a critical challenge in machine learning, with significant implications for fields ranging from molecular discovery and circuit design to cybersecurity. Despite its importance, existing graph generative models face substantial hurdles regarding scalability and the ability to produce novel structures. Methods based on diffusion often incur high computational costs due to full-adjacency operations and extensive denoising processes, while many autoregressive and hybrid architectures suffer from at least quadratic complexity. Furthermore, these models frequently fail to generalize beyond their training data, merely imitating existing graphs rather than creating new ones.

To overcome these limitations, we introduce a lightweight autoregressive framework. This approach employs a structure-guided topological ordering to transform graphs into standardized edge sequences, facilitating near log-linear generation speeds. Additionally, it utilizes a two-phase training regimen that merges exploration-focused augmentation with iterative refinement. This strategy is designed to mitigate overfitting and encourage controlled novelty.

Our experimental results across both molecular and non-molecular benchmarks demonstrate that this method successfully enhances novelty without compromising validity or uniqueness. The framework is versatile, supporting causal sequence backbones such as LSTMs and Mamba-style models. Moreover, the use of large-memory accelerators allows for the processing of longer graph sequences, surpassing the constraints typically imposed by standard GPU capabilities.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC