Challenge timeline
- NeurIPS 2026 competitions announced
The competition track publishes its accepted list. Nothing is open yet: the site, the portal and the board follow a fortnight later.
- P1Site and submission portal live; validation leaderboard opens
Everything opens at once: the site, the submission portal, and the validation leaderboard. Submissions for the validation split are scored and ranked from the first day: the validation answers themselves are not distributed, a submission comes back as a score. The daily submission quota is loosest here, while entrants are wiring their pipelines up.
- P2Starter kit and reference baselines released
Data loaders, the exact metric implementations, reference baselines, example submissions and executed tutorials: the public baselines repo is the reference implementation. This is the point where there is a published number to beat, and the daily quota tightens to match.
- P3Test phase starts, validation data released, test leaderboard opens
The point where the two halves swap. The validation data is released for every task (it can no longer be used to probe the board, because ranking moves to the hidden test split) so validation becomes something to develop against with the ground truth in hand. The test inputs are released without their labels, because a prediction needs something to predict from; the test answers never are. A test submission comes back as a score. Up to two official submissions per board are allowed for the whole phase, and the test leaderboard opens.
- Final submissions due
The last moment a submission can be made or a final entry nominated. Nothing after this counts, in either track.
- Final evaluationOfficial evaluation starts
Scoring runs on the hidden test sets across both tracks, and Agent Team entries are checked against the evidence attached to them.
- Winners announced at NeurIPS
Final rankings use the hidden test set, not the development leaderboard.
This page is the authoritative schedule and is updated whenever a date moves. A deadline may be extended; one is never brought forward without notice. Datasets and exact metric formulations can still be corrected. the evaluation page tracks the metric panel as it stands today.