BCIwiki (bciwiki.com) — Adaptive decoders that help brain-computer interface users perform better in the short term also fundamentally reshape how the motor cortex encodes task information over days, a team from the University of Washington and the Universite de Montreal reported in Nature Communications on September 15, 2026. In two rhesus macaques practicing a 2D cursor-reaching task, fewer neurons were needed to reach 80% task classification accuracy late in learning compared with early sessions (p = 9.8 x 10^-4). Fixed Wiener-filter decoders produced no such compaction.
The researchers recorded from 128-channel microwire electrode arrays implanted in motor and premotor cortices. Monkey J used a SmoothBatch adaptive decoder across seven learning series of at least four days each; monkey S used a ReFIT decoder across three series. At the population level, task information became largely confined to a small number of neural modes accounting for a minor fraction of total population variance. A neural network model suggested that adaptive decoders directly contribute to this compaction, which the authors propose may stem from altered error signals used in credit assignment — the brain's process for attributing outcomes to specific neural patterns. The authors argue the finding has practical implications for real-world BCI decoder design: assistive algorithms must be evaluated not only for immediate performance but also for their long-term effects on neural plasticity. The study was conducted in non-human primates and has not been validated in humans.