Neural decoders turn brain activity into intentions or predictions, and a recent advance was to tokenize the data at the level of individual spikes, which lets a model pretrain across many recording sessions. The catch: those spike models have needed labeled behavior to train on, and labels are scarce.

MOJO loosens that requirement by training on two objectives at once, a self-supervised masked-autoencoder task that just reconstructs hidden spikes alongside the usual supervised one. Across monkey motor-cortex reaching and multi-region mouse recordings, this beats supervised-only training, and the gap widens when labels are thin, exactly the few-shot situation a new session presents. The unsupervised half also yields representations that are better at classifying brain regions and predicting spike statistics without being trained for either. It even carries over to human electrocorticography during speech, matching purpose-built neuro-foundation models.

Encouraging breadth across species and tasks, though the paper is where the caveats and numbers sit.