Neuro Symbolic Graph Generative Modeling

Chuqin Geng, Li Zhang, Zhaoyue Wang, Haolin Ye, Ziyu Zhao, Xujie Si

Keywords: Graph Generative Models, Neuro-symbolic methods, SMT Solvers, Interpretability, Molecular Design

Abstract: We challenge the prevailing deep generative paradigm for graphs, exemplified by diffusion models, which is computationally intensive, lacks formal guarantees, and offers little user control. We introduce Neuro-Symbolic Graph Generative Modeling (NSGGM), which recasts graph generation as sequence modeling plus constraint satisfaction. NSGGM learns a vocabulary of subgraph tokens and uses an autoregressive sampler to propose token sequences, which an SMT solver then assembles into valid graphs by enforcing learned structural rules and user-defined constraints. This hybrid design avoids costly iterative refinement while providing correctness by construction and interpretable control. Across molecular and general graph benchmarks, NSGGM achieves state-of-the-art quality with fine-grained, user-steerable control, which current methods lack. Thus, NSGGM offers a practical path to trustworthy, targeted graph synthesis with broad applicability.

Primary Area: generative models

Submission Number: 2186