Reproduce “The 2026 PNPL Competition: Word Classification and Efficient Cross-Subject Generalisation in LibriBrain100” on OpenResearch

About this paper

Francesco Mantegna, Gereon Elvers, Dulhan Jayalath, Gilad Landau, Tasha Kim, Miran Özdogan, Luisa Kurth, Teyun Kwon and 13 more

The 2026 PNPL competition introduces LibriBrain100, an expanded MEG dataset, to benchmark non-invasive word classification and cross-subject generalization for speech Brain-Computer Interfaces. Reference models achieved a Top-10 Balanced Accuracy of 73.23% for within-subject decoding with extensive data and 42.96% for cross-subject generalization with limited fine-tuning.

Key insights

• The expanded LibriBrain100 dataset, with its unprecedented within-subject depth and multi-subject coverage, is foundational for advancing both high-performance single-subject decoding and robust cross-subject generalization in non-invasive BCIs.
• Standardizing word classification with specific vocabularies and metrics enables direct performance comparisons between non-invasive systems and, for the first time, against established invasive BCI systems.
• Self-supervised foundation models like MEG-XL demonstrate superior cross-subject generalization with limited fine-tuning data, highlighting their potential for clinical translation where rapid calibration to new users is essential.

Summary by alphaXiv.