Benchmark
Interactive skill acquisition. An agent is dropped into a grid-world game with no instructions and must work out the rules, the goal and the controls by playing, then clear every level.
No verified results recorded yet.
Agents act through a fixed action interface and cannot read the environment's source. Scores are reported as Relative Human Action Efficiency, which weights each level by how many actions a human baseline needed, with 100 as the ceiling. Entries on the public game set are recorded on ARC Prize scorecards; a held-out set exists separately.
A result on the 25 public games is not a result on the held-out set, and papers that saturate the public split often do not report the private one. Efficiency scores also depend on the harness and the compute budget behind the agent, so a number without its backbone model and action count says little.
A frozen Claude Opus 5 cleared every public ARC-AGI-3 game using 44% of the actions a human needed
The margin over the closest entry on the leaderboard is a single point. The number that moved is the action count: 7,518, against a human baseline of 17,135.
Privacy preferences
We use necessary cookies to run this site. With your consent we also use advertising cookies. You can change your choice at any time. Cookie Policy · Privacy Policy
