The 2026 PNPL Competition: Word Classification and Efficient Cross-Subject Generalisation in LibriBrain100

The 2026 PNPL Competition. Word Classification and Efficient Cross-Subject Generalisation in LibriBrain100. By Francesco Mantegna and twenty others, from the University of Oxford, Maastricht University, and three other institutions. Published on the third of September. Its code repository is worked out: about one hundred and twenty source files, with a readme and a licence.

What now exists is LibriBrain100, an extended version of the LibriBrain dataset, together with the 2026 PNPL competition built on top of it. LibriBrain was already the largest within-subject MEG dataset recorded at the time, holding roughly fifty hours of data for a single subject. That scale is what made strong decoding possible. It is also what limited the questions that could be asked, because a dataset with one subject in it cannot answer a question about new users.

The extension moves in two directions at once. Thirty-two additional subjects have been added, at roughly forty minutes of recording each, and the within-subject data has grown to around eighty hours. The dataset is therefore both deeper and wider than it was. The competition built on it advances the curriculum to word classification, and splits into two tracks that use those two directions separately. The Deep track and the Broad track ask different questions of the same recordings.

The lineage is explicit. The 2025 competition set out to launch a multi-year curriculum for non-invasive speech decoding, designed to progress from foundational tasks toward the linguistic complexity a practical interface would require, and it opened with speech detection and phoneme classification. Winning submissions reached F1-macro scores of ninety-five point six percent and seventy-three point six percent on those two tasks. Those results are why the curriculum advances rather than repeats: the foundational tasks were answered well enough to move on from.

The two tracks serve different users of the data. The Deep track targets within-subject word classification at scale, aiming at the best possible performance. The Broad track targets cross-subject generalisation, and does it by progressively reducing the amount of subject-specific fine-tuning data available, from around forty minutes to twenty and then to ten. That last figure is chosen rather than arbitrary. Ten minutes falls within a clinically feasible range, which is the constraint a real interface has to meet.

The reason the second direction matters is stated plainly. Within-subject scale drives strong decoding performance, but a practical brain-computer interface must generalise to new users from minutes of data rather than hours. Shrinking the fine-tuning budget until it becomes clinically feasible is a step toward a non-invasive interface capable of restoring communication to people living with profound paralysis.

The 2026 PNPL Competition. Word Classification and Efficient Cross-Subject Generalisation in LibriBrain100. By Francesco Mantegna and twenty others, from the University of Oxford, Maastricht University, and three other institutions. Published on the third of September. This is the long version: the whole paper rather than the abstract.

Invasive speech decoding has moved extraordinarily fast. Since the first demonstration of connected speech decoded from surgically implanted electrodes in a paralysed person, vocabularies have grown from fifty words to over a hundred and twenty-five thousand, and word error rates have fallen to two and a half percent. That is better than the five to ten percent usually reported for human transcription on standard speech recognition benchmarks. And yet the paper calls non-invasive decoding the real prize, for reasons that have nothing to do with accuracy. Brain surgery carries inherent risks. Per-patient data collection does not scale. And the people who would benefit most are often those least able to undergo elective neurosurgery.

Magnetoencephalography is the non-invasive modality with the best case. It offers millisecond temporal resolution and spatial precision typically around five to ten millimetres, which puts it in the range of the invasive modalities without the surgical risk. Real advances have been made with it. But progress has been harder to interpret and harder to compare across studies than in the invasive setting, and the authors are specific about why. What has been missing is not better models. It is shared infrastructure: common datasets, fixed evaluation splits, standard metrics, and baseline implementations that let a community build cumulatively rather than each group starting over.

The competition series exists to supply that. The first edition deliberately chose easy tasks, speech detection and phoneme classification, to lower the barrier for machine learning researchers without the domain knowledge to navigate brain data confidently. It shipped a library for automatic data download, standardised splits, reference models, a public leaderboard and tutorial code. Winning submissions reached F1-macro scores of ninety-five point six and seventy-three point six percent. The uptake is the part that justifies a sequel: a hundred and fifty-five registered teams from fifteen countries, over six thousand submissions, and by the following spring the underlying dataset had passed sixteen thousand downloads.

The dataset makes the two-track design possible by being unusual in two directions at once. LibriBrain100 adds thirty-two subjects at roughly forty minutes each, and deepens the single-subject recording to around eighty hours. Depth supports the Deep track, which pushes within-subject word classification as far as it will go. Breadth supports the Broad track, which tests cross-subject generalisation by progressively reducing the subject-specific fine-tuning data available, from about forty minutes to twenty and then to ten. That descending staircase is the whole scientific content of the second track. A practical interface must generalise to a new user from minutes of data rather than hours, and ten minutes is chosen because it falls within a clinically feasible range.

What this paper is, is a competition announcement, so there are no results yet: the tracks describe what will be measured, not what was found. Two scoping decisions stand out. The clinical feasibility of ten minutes is an argument about what a patient could tolerate, not a demonstration that ten minutes suffices; that is precisely what the track asks entrants to find out. And the authors explicitly exclude typing-based decoding from the speech interface problem, even though such models report interesting sentence-decoding results, on the grounds that there is no clear path for a typing-based method to serve patients unable to move their fingers.

The through-line is that this is infrastructure work presented as a competition. The dataset, the splits, the library, the leaderboard, the reported vocabulary, the metric that survives a change of vocabulary: none of it is a model, and all of it is what makes the models comparable. The stated destination is a non-invasive interface capable of restoring communication to people living with profound paralysis, and the argument being made is that the field gets there by building common ground first.