WARC is a 35M-parameter action-conditioned vision Transformer: it takes the last three frames of 64×64, 16-color game screens and actions as input, and predicts the next frame cell by cell; each time the agent observes a state change, it performs 20 steps of gradient updates on all parameters using the most recent transition. The authors claim that after pre-training on human gameplay recordings, for unseen games, four out of five mechanisms achieve at least 80% prediction accuracy on changed cells. The repository provides four checkpoints, three split by held-out games, and one trained on all 25 games.
Why it matters Demonstrates how small models can quickly master unfamiliar game mechanisms through test-time training; weights are released under CC0, allowing direct reuse in research.
Hug face likes 2 · Hug face upvotes 2 · Hug face downloads 0
License: Commercial use allowed · ComfyUI: Not available · VRAM: Not specified
Reasoning
- CC0 license, code is public on GitHub, each checkpoint includes training configuration, per-epoch validation accuracy, and SHA-256 manifest
Usage
- Clone the GitHub repository, install torch 2.7.0 and dependencies according to the README, download checkpoints using hf download third-intelligence/WARC, then call build_from_checkpoint to predict the next frame
Risks
- Only applicable to ARC-AGI-3 style grid games, intended for research purposes
- Requires a CUDA GPU, model card does not specify VRAM requirements