XCAVATE: A CROSS-MODEL PLATFORM FOR MAPPING CONCEPTS ONTO THE HUMAN BRAIN

Understanding the large-scale organization of human concepts remains challenging because cognitive, affective, clinical, and phenomenological domains lack a shared quantitative reference frame. Here we introduce XCAVATE, a data-driven screening and aggregation method that uses cross-model replicated inference from large language models (LLMs) to map concepts onto a standardized human brain map. XCAVATE estimates stable activation profiles across 80 anatomically defined brain regions for 260+ diverse concepts spanning neuroscience, psychology, psychiatry, learning theory, and contemplative traditions, collapsing semantically related terms to emphasize shared network structure rather than fine-grained taxonomic distinctions. Using 50 frontier-class LLMs with multiple stochastic replicates per model, we identify robust, cross-model activation patterns aligning with ground-truth anatomical regions and validated neuroscience concept:brain associations, with sub-groups of concepts mapping distinctively onto the brain’s canonical triple-network organization: a midline/default-mode network associated with awareness, and identity, a frontoparietal control network associated with attention and metacognitive access, and salience networks associated with interoceptive functions. Together, these results show that LLMs encode human concepts in a low-dimensional geometry that converges on established large-scale brain networks, providing a unified empirical framework for relating cognitive, emotional, clinical, and phenomenological constructs.