Proceedings of the Annual Meeting of the Cognitive Science Society

A central challenge in cognitive neuroscience is identifying the algorithmic principles that allow the human brain to transform linear linguistic input into hierarchical macro-syntactic structures. We propose the Geometric Resonance Hypothesis, postulating that the Left Angular Gyrus (L-AG) operates on a representational geometry that is topologically isomorphic to the deep derivational manifolds emergent in Large Language Models (LLMs). Utilizing fMRI data from the Natural Stories Corpus and the BERT transformer architecture, we track the alignment between neural activity and the model's computational hierarchy. Residual Representational Similarity Analysis (RSA) reveals a distinct phase transition: neural alignment remains minimal in lexical layers but peaks significantly in the deepest integration layers. Quantitative manifold analysis demonstrates that this resonance is driven by the intrinsic structural complexity of the representational space. Furthermore, a causal mediation analysis establishes that the relationship between syntactic complexity by Dependency Locality Theory scores and L-AG BOLD responses is significantly mediated by the model's macro-syntactic manifold. These results suggest a biological convergence, where disparate substrates, biological circuits and artificial networks, converge upon shared geometric solutions to resolve the complexity of human language.