Skip to content
arXiv Game World Models & Generation· Ku Onoda·· 13 d agoPickAI score54

How many interactions does 35M-parameter WARC need to learn unseen game mechanics on ARC-AGI-3

How Many Exposures Does a World Model Need? Learning Unseen Game Mechanics at Test Time

AI summary

The authors introduce MechEval on ARC-AGI-3, labeling 69 recurring mechanics across 20 games, and build WARC, a 35M-parameter vision Transformer world model, to examine how many interactions it needs to predict the effects of mechanics in a new environment.

Why it matters

The paper offers a quantifiable evaluation protocol, so readers can judge how many interactions a world model needs to learn new mechanics in an unfamiliar game.

Source: arXiv Game World Models & Generation · arxiv.org