This is a clean, open-source, single-file implementation of the model architecture from the "ARC Is a Vision Problem" paper. It fully features the paper's Canvas mechanism and a fast transformer with grouped-query attention + RMSNorm.
| Key Contribution | Description |
|---|---|
| Vision-centric formulation | Reformulates ARC as an image-to-image translation problem rather than a language-oriented task |
| Canvas representation | Introduces a canvas-based representation that enables processing ARC tasks as natural images |
| Standard vision architecture | Demonstrates that a vanilla Vision Transformer can effectively solve ARC tasks when trained from scratch |
| Test-time training | Employs test-time training for generalization to unseen tasks |
| Strong performance | Achieves 60.4% accuracy on ARC-1, outperforming existing from-scratch methods and competing with leading LLMs |
pip install -r requirements.txtimport torch
from varc.main import VisionARCModel, ARCLoss
# Example usage
if __name__ == "__main__":
# Initialize model with grouped query attention
model = VisionARCModel(
num_colors=10,
canvas_size=64,
patch_size=2,
d_model=512,
n_layers=10,
n_query_heads=8, # Number of query heads
n_kv_heads=2, # Number of key-value heads (shared across query groups)
mlp_dim=512,
dropout=0.1,
num_tasks=400,
)
print(f"Model parameters: {sum(p.numel() for p in model.parameters()) / 1e6:.1f}M")
# Example forward pass
batch_size = 4
dummy_canvas = torch.randint(0, 10, (batch_size, 64, 64))
task_id = 0
logits = model(dummy_canvas, task_id)
print(f"Output shape: {logits.shape}") # (B, 64, 64, 12)
# Example loss computation
dummy_target = torch.randint(0, 10, (batch_size, 64, 64))
loss_fn = ARCLoss(bg_token=10)
loss = loss_fn(logits, dummy_target)
print(f"Loss: {loss.item():.4f}")See example.py for a complete usage example.
For comprehensive API reference documentation, see VisionARCModel Reference.
The VARC framework consists of several key components:
- Canvas: Handles placement of ARC grids on a fixed-size canvas with scaling and translation
- Vision Transformer: Standard ViT architecture adapted for image-to-image translation
- Separable Positional Embeddings: 2D separable positional embeddings for spatial awareness
- Task Conditioning: Task-specific tokens for multi-task learning
- Data Augmentation: Geometric and color-based augmentations for improved generalization
- ARC-1 Accuracy: 60.4%
- Training: From scratch on ARC data only
- Generalization: Test-time training for unseen tasks
If you use this code in your research, please cite:
@article{hu2025arc,
title={ARC Is a Vision Problem!},
author={Hu, Keya and Cy, Ali and Qiu, Linlu and Ding, Xiaoman Delores and Wang, Runqian and Zhu, Yeyin Eva and Andreas, Jacob and He, Kaiming},
journal={arXiv preprint arXiv:2511.14761},
year={2025}
}See LICENSE file for details.
This is a community implementation of the VARC framework. For the official implementation and project webpage, please refer to the original repository.