How Many Exposures Does a World Model Need? Learning Unseen Game Mechanics at Test Time
With test-time updates, WARC predicted four in five unseen game mechanics at at least 80% accuracy on the cells they changed.
The 35-million-parameter model updates all its parameters after each game transition. Researchers evaluated it using MechEval, which labels 69 recurring mechanics across 20 ARC-AGI-3 games. Half of the successfully learned mechanics reached the accuracy threshold within about 80 observed occurrences; without updates, only one in seven reached it at all. Paired with a fixed graph explorer, WARC raised the explorer’s public-game efficiency score from 0.15 to 0.46; the team’s hidden-game leaderboard entry scored 0.20.
HF Daily Papers' note
The 35-million-parameter model updates all its parameters after each game transition. Researchers evaluated it using MechEval, which labels 69 recurring mechanics across 20 ARC-AGI-3 games. Half of the successfully learned mechanics reached the accuracy threshold within about 80 observed occurrences; without updates, only one in seven reached it at all. Paired with a fixed graph explorer, WARC raised the explorer’s public-game efficiency score from 0.15 to 0.46; the team’s hidden-game leaderboard entry scored 0.20.
HF Daily Papers' note
score 5