Neuro-native paradigms of brain foundation models: evolution, taxonomy, and perspectives

The deep integration of artificial intelligence and neuroscience has catalyzed the development of brain foundation models designed to extract transferable neural representations from vast and unlabeled neurophysiological datasets. Initial first-generation models largely relied on isomorphic transfer by migrating architectures from large language models and vision transformers to the neural signal domain. Although these approaches demonstrated the value of scale, evidence across the literature also reveals recurring mismatches between general-purpose architectural assumptions and the spatial, temporal, and acquisition properties of neurophysiological recordings. Yet the field lacks a unified account of how models have moved from general-purpose architectural transfer to designs that explicitly encode neurophysiological structure, and why this transition matters for cross-subject and cross-device generalization. This review addresses this need by tracing the evolution from 2021 to 2025 of EEG-centered neuroelectrophysiological foundation models from isomorphic adaptation toward second-generation neuro-native architectures. Unlike simple parameter expansion, these next-generation architectures are characterized by the explicit incorporation of neuro-inductive biases. We organize these models into four core neuro-native paradigms based on their design principles, encompassing topology awareness using coordinate manifold mapping, brain-like efficiency employing linear state space models, multi-scale hierarchy mirroring recursive organization, and semantic fusion bridging the neuro-semantic gap. Taken together, these developments suggest that neurophysiological priors can improve transfer and computational efficiency when they match the task and acquisition setting. Their broader clinical and causal value remains to be established through controlled, cross-subject, cross-device, and intervention-aware validation.