Neural Proposals, Symbolic Guarantees: Neuro-Symbolic Graph Generative Modeling
NeuroSymbolic Graph Generative Modeling reapproaches molecule generation by treating it as a scaffold and interaction learning task. It uses a neural model to propose structural scaffolds and refinement signals, while an SMT solver deterministically builds the final graph. This method offers strong generative performance with explicit, verifiable control over chemical validity through hard constraints.