BCIwiki (bciwiki.com) — A Tsinghua University team has proposed LAND (Latent Aligned Neural-behavioral Dynamics), a framework that uses flow matching to align neural and behavioral dynamics in latent space, aiming to solve one of the persistent problems in invasive brain-computer interfaces (BCIs): decoders that break down when conditions change. The work was posted as a preprint on bioRxiv on August 24, 2026, and has not been peer reviewed.
Invasive BCI decoders are typically trained on paired neural-behavioral data collected under constrained calibration conditions. When movement speed, the limb being controlled, or the subject itself changes, decoding accuracy tends to drop sharply. LAND addresses this by learning a neural-to-behavioral transport map through flow matching and using behavioral-dynamics priors drawn from unpaired kinematic data to regularize neural manifold geometry, encouraging representations that transfer across domains.
The team evaluated LAND on three datasets: synthetic neural data, epidural BCI recordings from a participant with tetraplegia, and multi-electrode array recordings from nonhuman primates. Across these settings, LAND achieved zero-shot generalization to out-of-distribution movement speeds and produced speed-modulated manifolds. With limited fine-tuning on target-domain data, cross-limb and cross-subject transfer performance also improved.
The authors are Yao R., Zheng J., Wang Y., Li W., Zou X., and Hong B. As a preprint, the findings await peer review. The evaluation did not include online closed-loop decoding scenarios.