Update 4/10: IEEE acceptance

Graph2Proof is the first framework that brings topology-aware methods to natural language theorem proving in LLMs. It converts mathematical proofs into directed acyclic graphs (DAGs) and uses a pre-trained Graph Attention Network (ProofGAT) to provide dense structural feedback during reinforcement learning. Additionally, an interpretability framework based on spectral graph theoretic methods shows for the first time that correct proofs carry measurable topological signatures: more compact structure, higher closeness centrality, more scale-free degree distributions, and greater structural resilience than incorrect proofs.