Human effort. Best = fewest published actions for a full-game win. Average = winning entries in the published top-10 human leaderboard, not all players. Counts are rounded to whole actions.
AI score. ARC Prize’s action-efficiency score, not completion percentage. Each cell uses one run. WIN and completed levels are shown separately; a zero score is different from no result.
Winning actions. Only full-game wins qualify. The gap compares that AI run with the human best for the same game. Failed or partial runs never count as short solutions.
Published configurations. “Best available” can choose different settings per game. Select a fixed setting beneath a model’s name for consistent comparisons. Win counts and filters use all runs allowed by that setting.
Dated snapshot of the public tutorial set. Best records are not proven optima. These results do not measure performance on the private competition games.