Hierarchical Model Trees for Semantic Summarization and Adaptive Inference

Learning systems must handle scale while remaining context-aware, adaptive, and interpretable. Symbolic methods are precise but rigid; neural models generalize but obscure structure. We propose a hybrid hierarchical system that organizes data into a dynamic tree of contextual models. Each node maintains both explicit (e.g., topic/cluster signals) and latent (neural) representations, adapting the hybridization locally. The system refines itself via border-sensitive updates and localized restructuring, avoiding costly global retraining. On Reddit and Cora, our approach outperforms strong baselines for classification and recommendation, with up to 47 % F1 and 150 % MRR improvements across epochs. The result is an efficient, interpretable framework that integrates symbolic and neural reasoning to build and maintain adaptive context at scale.