BCIwiki (bciwiki.com) — University of Oxford researchers Miran Özdogan, Gereon Elvers, Michael Bronstein and Oiwi Parker Jones compared ways to combine repeated neural trials. Averaging classification scores during training led in six of seven settings and exceeded the 2025 PNPL competition winner by 17.4 percentage points in phoneme-classification balanced accuracy. The preprint was uploaded to arXiv on October 19, 2026 and has not been peer reviewed.
Repeating a stimulus and combining trials is one way to address low signal-to-noise ratios in non-invasive recordings. The team compared six strategies across three datasets, four tasks and two group sizes, including training-time logit averaging and three attention-based variants.
Logits are classification scores before conversion to probabilities. Averaging them during training differs from averaging input brain signals or combining prediction probabilities after training.
The authors also discuss a theoretical connection to joint ensemble training. The 17.4-point advantage concerns repeated-trial aggregation; single-trial decoding and practical communication speed require separate assessment.
The full paper specifies that the main phoneme comparison uses LibriBrain MEG recordings from one participant listening to English audiobooks, aggregating 100 trials per phoneme group. The 17.4-point gap is a model re-evaluation on the public test split: balanced accuracies of 64.1% and 46.7%, averaged over five random seeds. Labels for the competition holdout are unavailable, so this is not a direct leaderboard comparison. The experiments use MEG; transfer to EEG has not been evaluated.