Four NeurIPS-track competitions across five years, with the same lead and platform. Below: what each prior round did and how the 2026 edition builds on it.
EEG Foundation Challenge, the immediate predecessor of the 2026 edition. NeurIPS competition track, sealed test, code submissions on Codabench.
First in the lineage. Cross-subject motor-imagery EEG decoding across heterogeneous BCI datasets. Established the team's pattern of code-submission scoring with a held-out evaluation set, and the use of Braindecode + MOABB as the canonical stack.
First time cross-subject EEG transfer ran as a NeurIPS competition, with code submissions on a held-out set and an open-source stack from day one.
EEG-based age regression on a multi-cohort benchmark. Stress-tested cross-site generalisation and introduced the dataset-aggregator pattern that NeuralBench inherits today.
Multi-cohort aggregation as a benchmark axis. The pattern that became NeuralBench's dataset registry.
Sleep-stage classification across PhysioNet, Sleep-EDF, and HMC sources. Wearable-grade data plus expert-scored hidden labels; the same recipe the 2026 Sleep track uses, scaled up to 1,000+ subjects and the Muse Sleep-Onset corpus.
Wearable EEG with expert-scored hidden labels. The Sleep-track recipe, now scaled.
The immediate predecessor. Cross-task EEG decoding with single-task and foundation-transfer arms. The 2026 challenge is its operational continuation: same lead, same Codabench infrastructure, same open-source stack, plus EMG as a parallel modality.
Same lead. Same platform. Same stack. EMG added as a parallel modality.
Three additions over the 2025 edition. Each is a deliberate extension of the same competition design, not a rewrite.
First time surface EMG (wristband hand pose, emg2pose corpus + 100 new evaluation users) sits inside the same sealed benchmark as EEG. The harness, splits, and audit are identical across modalities, so EEG and EMG results are directly comparable — a setup the 2025 edition could not pose.
The four tracks evaluate task-specific decoders on the same data and metrics, so a per-track score is attributable to the model rather than the protocol.
Each task-specific track isolates one primary shift (cross-stimulus, cross-session, cross-subject, cross-user) so post-competition analysis can be shift-specific rather than leaderboard-only.
The 2026 edition keeps every load-bearing piece of the 2025 round. Lead, platform, and open-source stack are unchanged.
Bruno Aristimunha leads both the 2025 and 2026 editions. Yneuro hosts both.
Codabench (Université Paris-Saclay, Sylvain Chevallier) hosts both. The 2026 challenge funds and contributes upstream a scaled AWS-backed deployment with Deloitte.
EEGLAB, MNE-Python, Braindecode, MOABB, EEGDash, NeuralSet, NeuralBench. The 2026 baselines reproduce in NeuralBench out of the box.
Five years of EEG decoding competitions, condensed into one four-track benchmark. Open-source, sealed test, NeurIPS-track audit.