Rules & FAQ.

Industry, academia, students, and independents are all eligible, and cross-institution teams are encouraged. Organisers and their direct team members may submit, but their entries appear as Organizers on the leaderboard and are ineligible for cash prizes.

Each participant submits through their own Codabench account and has an individual submission quota. For team entries, those submissions are attributed to the participant's Codabench organization and ranked as team submissions.

Warm-up (Sep 21 to Oct 25, 2026): 5 submissions per participant per day. Sealed final (Oct 28 to Nov 21, 2026, AoE): 1 per participant per day, so the prize roster reflects deliberate iteration rather than lottery search. The final NeurIPS ranking uses sealed submissions only.

Training data is suggested to stay within the track, but pre-training on any publicly available, redistributable dataset is allowed. The sealed test split is never allowed, and closed clinical datasets are not allowed. Original dataset and model licenses remain in force. Declare every external corpus, and a compute estimate, in the method description that ships with your final submission. The audit cross-checks it against a re-run of your submitted training pipeline and declared configuration.

Train on whatever you have. The scoring container must complete a full test pass in under 60 minutes on one A100 GPU, so that audit cost stays bounded and per-team runtime stays comparable. AWS, the compute partner, provides complimentary instance hours to finalists.

Codabench also limits each profile to 15 GB of submission storage shared across all tracks and uploads. This makes 15 GB the theoretical maximum for one submission, but your practical limit is the space remaining in your profile. If needed, delete unused archived submissions under Profile → Resources. A submission displayed on a leaderboard cannot be deleted.

Any attempt to cheat, including but not limited to using closed clinical data, accessing the sealed test split, or modifying the scoring harness, results in immediate disqualification from the competition. This decision is made by the organising committee and is final.

After the sealed final closes, the top-3 of every track go through a reproducibility audit: we re-run your training pipeline from the committed config and re-score the resulting weights against the sealed split. Scores within ±2σ of your submitted number stay on the prize roster. Scores outside the tolerance drop off the prize roster but remain on the public board for context. The organising committee chairs the audit.

Participants retain the intellectual property rights they hold in their submitted code, models, and related materials. Finalists must provide the materials required for the audit and grant the auditing committee permission to access and run them solely to verify the submission and detect cheating. This limited audit permission does not allow the organisers to publish, share, modify, commercialise, or otherwise use those materials outside the audit. Open-source release after the competition is encouraged, but remains entirely the participant's choice.

Top-ranked teams that pass the audit and submit a method description, training and inference code, and pre-training disclosures are announced during the NeurIPS workshop. This opportunity is offered to teams with complete, reproducible artifacts.