Abstract Draft

Brain foundation models increasingly cover stimulus-to-fMRI prediction and unimodal or paired-modality pretraining, but progress is difficult to compare under strict leakage controls. NeuroTwin introduces a neural translation benchmark and NeuroTwin NFC, a Neural Field Compiler that treats heterogeneous recordings as partial observations of a latent neural field before compiling modality-specific outputs. The benchmark evaluates missing-modality reconstruction, future-state forecasting, and few-shot adaptation under held-out subject, site, and dataset splits. Pair-Operator is evaluated only as an ablation. Synthetic and debug results are plumbing checks; scientific claims require real-data evaluation against strong baselines.