LAND: Latent Aligned Neural-Behavioral Dynamics via Flow Matching forGeneralizable Movement Decoding

Generalizable movement decoding remains a central challenge for invasive brain--computer interfaces (BCIs), as decoders trained under limited calibration conditions often fail to generalize to unseen movement speeds, limbs, and subjects. Existing decoding methods are typically trained on paired data collected under restricted conditions. How to incorporate behavioral structure from unpaired data for robust out-of-distribution (OOD) decoding therefore remains unresolved. To address this, we propose LAND (Latent Aligned Neural-behavioral Dynamics), a framework that aligns latent neural and behavioral dynamics through flow matching. By learning a neural-to-behavioral transport map and using behavioral-dynamics priors from unpaired data to encourage structured neural manifolds, LAND regularizes representation geometry to promote cross-domain generalization. We evaluate LAND on synthetic neural data, epidural BCI recordings from a tetraplegia participant, and multi-electrode array (MEA) recordings from nonhuman primates (NHPs). Across these settings, LAND improves zero-shot generalization to OOD movement speeds and yields speed-modulated manifolds. With limited target-domain fine-tuning, it further improves transfer across limbs and subjects. These results support flow-based neural--behavioral alignment with unpaired kinematic priors as an approach for learning transferable neural representations and robust movement decoding across behavioral and recording domains.