Unifying Generative and Classification-Based Relation Extraction via MCTS for nextG Protocol Formal Verification

Formal verification of domain-specific protocols, such as 5G RRC and emerging defense communication standards, is essential for ensuring system reliability. However, automated extraction of formal relationships remains challenging due to limited annotations, evolving terminology, and out-of-vocabulary (OOV) conditions. We propose a unified framework that integrates classification and generative models using Monte Carlo Tree Search (MCTS) as a search controller. MCTS navigates a hybrid search tree where classification (CAL) and generation (REBEL) correspond to orthogonal expansion paths, with a learned reward model guiding which reasoning strategy to apply at each step. This unified search structure enables the framework to balance the contextual flexibility of generative expansions with the precision of classification-based inferences. A reinforcement learning-based reward model estimates the utility of candidate actions during simulation and guides value propagation throughout the MCTS tree, enabling robust extraction under low-resource and OOV conditions. Experimental results on the 5G RRC dataset show that our MCTS-based method achieves 97.2% accuracy in in-vocabulary settings and 91.2% in OOV scenarios outperforming REBEL and CAL. These results highlight the potential of reward-guided hybrid search to improve scalable, interpretable formal verification pipelines in mission-critical environments.