Foundational Human Modeling: A roadmap for brain-grounded multi-modal AI

Foundation agents excel at language, perception, and tool-use, yet remain brittle across domains and time: their success rates can exhibit ``half-lives'' under distribution shift, and their reasoning is often forced through language tokens \citep{foundationagents2025,jagged_frontier_2023,agent_success_half_life_2025,embers_autoregression_2024}. This roadmap argues for \emph{Foundational Human Modeling (FHM)}: scaling generative brain models (EEG/MEG/ECoG/sEEG/fMRI) and integrating \emph{brain tokens} into multimodal token streams to provide privileged internal signals for grounding, regularization, and human-like cognition. We propose: (i) scaling next-brain-token prediction toward stable long-horizon neural rollouts \citep{csaky2026scalingnextbraintoken,csaky2024gpt2meg,meg_gpt_2025}; (ii) unifying heterogeneous brain modalities via shared geometry and adapters \citep{xiao2025brainomni,scaling_laws_neural_data_2024}; and (iii) treating cognition as an interleaved typed token stream in which brain, vision, language, and action tokens can be produced in a dynamic ordering