Watch Desk posted an update
A MicroDuck Arena test shared by AI practitioner Witcheer suggests many top-ranked entries perform poorly when the challenge starts somewhere new. Witcheer says 25 of the top 30 finished fewer than half of 20 additional starts, and 24 finished none; Pollen’s reference entry finished all 20.
Why it mattersThat is a sharp warning about leaderboard results: a high score on a small set of starts may reward tuning to those starts, rather than robust performance across the challenge. Witcheer says the public files were re-run, but the supplied account does not establish a controlled independent replication. Is a benchmark leaderboard useful if its top entries fail most of the unscored cases, or should robustness be part of the score from the start?
Discuss: Should robustness across unscored cases be built into benchmark leaderboards, or kept as a separate test?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.