PNPL 2026 contest advances non-invasive BCI to word decoding

The 2026 PNPL competition asks models to classify whole words from magnetoencephalography recordings, and on one of its two tracks to do so after only around 10 minutes of fine-tuning on a new subject.

The previous year's contest set the floor. "Winning submissions reached F1-macro scores of 95.6% and 73.6% on the respective tasks, highly significant advances," the paper reports for speech detection and phoneme classification. Those scores were built on LibriBrain, which the authors describe as "the largest within-subject MEG dataset recorded at the time with ∼50 hours of data for one subject."

The new release, LibriBrain100, adds 32 more subjects at roughly 40 minutes each and pushes the primary subject to about 80 hours.

The word-classification jump comes with a split. The Deep track chases within-subject peak performance; the Broad track, in the authors' framing, "targets cross-subject generalisation, progressively reducing the amount of subject-specific fine-tuning data from ∼40 to ∼20 to ∼10 minutes." The organisers describe that final step as "a duration that falls within a clinically feasible range" for a BCI aimed at "restoring communication to people living with profound paralysis."

The abstract lists no baseline word-classification score, no target vocabulary size, and no demographic detail on the 32 new subjects.