This is the project webpage for the paper ARC Is a Vision Problem!

We formulate ARC with a vision paradigm, casting it as an image-to-image translation task.

🌟🌟Explore the full gallery on ARC-1 VARC ↗🌟🌟

🌟🌟Explore the full gallery on ARC-1 VARC-ensemble ↗🌟🌟

🌟🌟Explore the full gallery on ARC-2 VARC ↗🌟🌟

🌟🌟Explore the full gallery on ARC-2 VARC-emsemble ↗🌟🌟

Attention heatmaps highlighting how VARC aligns pixels between inputs and outputs.
🌟🌟Try interactive attention demos ↗🌟🌟

t-SNE of task embeddings, on the 400 task tokens learned from the ARC-1 training set. Each point represents a single task.

🌟🌟Try interactive task token demos ↗🌟🌟

conda create -n visarc python==3.10
conda activate visarc
pip install -r requirements.txt
hf auth login
wandb login

https://huggingface.co/VisionARC/offline_train_ViT/tree/main

https://huggingface.co/VisionARC/offline_train_Unet/tree/main

These checkpoints are from offline training, so they still need to be evaluated with TTT.

mkdir saves/
mkdir saves/offline_train_ViT
mkdir saves/offline_train_Unet
hf download VisionARC/offline_train_ViT --local-dir saves/offline_train_ViT
hf download VisionARC/offline_train_Unet --local-dir saves/offline_train_Unet

ARC-1 and ARC-2 test-time training (TTT) use the same checkpoint from ARC-1 training dataset + RE-ARC dataset

These predictions are made after TTT on each task.

hf download VisionARC/VARC_predictions --local-dir . --repo-type dataset
unzip VARC_predictions.zip

After downloading the predictions, run analysis to get results and generate HTML visualizations.

ARC-1 results:

bash script/analysis/arc_1_vit.sh
bash script/analysis/arc_1_ensemble.sh

The HTML visualizations will be at arc_agi_1_vit.html and arc_agi_1_ensemble.html.

ARC-2 results:

bash script/analysis/arc_2_vit.sh
bash script/analysis/arc_2_ensemble.sh

The HTML visualizations will be at arc_agi_2_vit.html and arc_agi_2_ensemble.html.

This also includes commands to sanity check the augmented versions.

# Build augmented data
python augment_data.py
# Run the following command
bash script/sanity_ARC1.sh
bash script/sanity_ARC2.sh

Train VARC-ViT-18M (5h 12m 42s on 8 x H200):

bash script/offline_train_VARC_ViT.sh 

Train VARC-Unet-55M (7h 21m 29s on 8 x H200):

bash script/offline_train_VARC_Unet.sh

Train VARC-ViT-18M with TTT:

bash script/test_time_training_VARC_ViT_ARC1.sh

Each run's result can range between [52, 56].

Train VARC-Unet-55M with TTT:

bash script/test_time_training_VARC_Unet_ARC1.sh

Each run's result can range between [47, 49].

Train VARC-ViT-18M with TTT:

bash script/test_time_training_VARC_ViT_ARC2.sh

Each run's result can range between [6, 10].

Train VARC-Unet-55M with TTT:

bash script/test_time_training_VARC_Unet_ARC2.sh

Each run's result can range between [3, 6].

Run analysis to get final results and generate HTML visualizations (same as with downloaded predictions) by modifying --output-root to the path you save your predictions (e.g., outputs/ARC_1_eval_ViT_attempt_0).

bash script/analysis/arc_1_vit.sh
bash script/analysis/arc_1_ensemble.sh

The HTML visualizations will be at arc_agi_1_vit.html and arc_agi_1_ensemble.html.

Below is our verified result using VARC-ViT-18M (no ensembling) on the ARC-2 private test set from the Kaggle competition, using a slightly smaller computation setting with 4 color permutations instead of 9.

Leaderboard: https://www.kaggle.com/competitions/arc-prize-2025/leaderboard

We use test pairs from the original ARC-1 training tasks as our validation.

If you find our method or models helpful, please kindly cite our paper :)

@misc{hu2025arcvisionproblem,
      title={{ARC} Is a Vision Problem!}, 
      author={Keya Hu and Ali Cy and Linlu Qiu and Xiaoman Delores Ding and Runqian Wang and Yeyin Eva Zhu and Jacob Andreas and Kaiming He},
      year={2025},
      eprint={2511.14761},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2511.14761}, 
}