WARC learns unseen game mechanics from limited test-time exposures
Original title: How Many Exposures Does a World Model Need? Learning Unseen Game Mechanics at Test Time
The 35M-parameter model predicts four in five held-out mechanics at 80% cell accuracy; half reach that within about 80 exposures.
The paper studies how many exposures a world model needs to predict mechanics in an unfamiliar environment, using ARC-AGI-3 and MechEval. MechEval labels 69 recurring mechanics across 20 games and scores changed cells. WARC, a 35M-parameter vision transformer pretrained on other games, updates all parameters after each transition. WARC-Agent triples a fixed explorer's efficiency from 0.15 to 0.46, mostly through test-time training, and scores 0.20 on hidden Kaggle games.
- Without parameter updates, only one in seven mechanics reaches that accuracy.
- Accuracy is retained through the end of the recordings.
- Closed-loop gains come mostly from test-time training.
- Products
- WARC-AgentMechEvalARC-AGI-3
- AI models
- WARC
From the original article
1 min read at huggingface.co
World models enable agents to plan and learn from imagined experience, but in an unfamiliar environment they must learn its mechanics from limited interaction. A mechanic is a rule for how an action or an interaction changes part of the environment. Each occurrence of a mechanic that the model observes is one exposure.
By Ku Onoda, Haruto Kawano, Shota Takashiro · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.