| --- |
| license: cc0-1.0 |
| tags: |
| - world-model |
| - arc-agi-3 |
| - test-time-training |
| - next-frame-prediction |
| - vision-transformer |
| - pytorch |
| --- |
| |
| # WARC |
|
|
| [](https://github.com/third-intelligence/WARC) |
| [](https://creativecommons.org/publicdomain/zero/1.0/) |
|
|
| WARC (World model for ARC-AGI-3) is a 35M-parameter action-conditioned vision transformer. It |
| predicts the next frame of a game, cell by cell, from the last three frames and |
| the action, and it keeps learning while an agent plays: after every observed |
| transition it updates all of its parameters by gradient descent. Pretrained on |
| human recordings of other games, it learns to predict four in five mechanics |
| of an unseen game with at least 80% accuracy on the cells they change and keeps |
| that accuracy to the end of the recordings. |
|
|
| ## Model details |
|
|
| | | | |
| |---|---| |
| | Input | The last three 64x64 frames of 16 colours, the action after the last frame, and the two actions between the frames | |
| | Output | A distribution over the 16 colours of every cell of the next frame | |
| | Architecture | 2x2 patches, 10 pre-norm transformer blocks of width 512 with rotary position embeddings, modulated by the action through adaptive layer normalization; a learned vector added to the token of the clicked cell | |
| | Parameters | 35M | |
| | Pretraining | Next-frame cross-entropy on human recordings of the public games; Muon and AdamW, 30 epochs | |
| | Test-time training | 20 gradient steps on the ten newest and twenty earlier transitions after every observed transition; all parameters are updated | |
| | License | CC0 1.0 | |
|
|
| ## Files |
|
|
| | File | Pretrained on | Held-out games | Validation exact-frame accuracy | |
| |---|---|---|---:| |
| | `fold1/checkpoint_best.pt` | 16 games | bp35 lf52 lp85 r11l re86 sb26 sc25 su15 wa30 | 85.7% | |
| | `fold2/checkpoint_best.pt` | 17 games | cd82 cn04 dc22 m0r0 sk48 tn36 tr87 tu93 | 79.2% | |
| | `fold3/checkpoint_best.pt` | 17 games | ar25 ft09 g50t ka59 ls20 s5i5 sp80 vc33 | 78.5% | |
| | `all25/checkpoint_best.pt` | all 25 games | none (one recording per game held out for validation) | 80.0% | |
|
|
| Each checkpoint (146 MB) holds the weights, the model configuration, and the |
| game split; `config.json` and `eval_history.jsonl` next to it record the |
| training configuration and the per-epoch validation accuracy. `manifest.json` |
| lists the SHA-256 of every checkpoint and the fold of every game. |
|
|
| ## Usage |
|
|
| Install the code from the [GitHub repository](https://github.com/third-intelligence/WARC) |
| and download the weights: |
|
|
| ```bash |
| git clone https://github.com/third-intelligence/WARC |
| cd WARC |
| python3.12 -m venv .venv && source .venv/bin/activate |
| pip install torch==2.7.0 torchvision==0.22.0 --index-url https://download.pytorch.org/whl/cu126 |
| pip install -r requirements.txt |
| hf download third-intelligence/WARC --local-dir models # the four checkpoints (pip install huggingface_hub) |
| ``` |
|
|
| Predict the next frame: |
|
|
| ```python |
| import sys |
| import torch |
| |
| sys.path[:0] = ["agent", "agent/VARC"] |
| from warc_model.model import build_from_checkpoint |
| |
| device = torch.device("cuda") |
| model, meta = build_from_checkpoint("models/all25/checkpoint_best.pt", device=device) |
| |
| # The last three frames, (1, 3, 64, 64) colours 0-15. Actions: 1-5 are the keys ACTION1-ACTION5, |
| # 6 is a click at column x and row y; x and y are -1 for a key. |
| frames = torch.randint(0, 16, (1, 3, 64, 64), device=device) # replace with frames of a game |
| action, x, y = (torch.tensor(v, device=device) for v in ([6], [10], [20])) |
| hist_action, hist_x, hist_y = (torch.tensor(v, device=device) for v in ([[1, 2]], [[-1, -1]], [[-1, -1]])) |
| |
| with torch.no_grad(): |
| logits = model(frames[:, -1], torch.zeros(1, dtype=torch.long, device=device), action, x, y, |
| hist_obs=frames, hist_action_ids=hist_action, hist_action_xs=hist_x, |
| hist_action_ys=hist_y)["ffp_logits"] # (1, 16, 64, 64) |
| next_frame = logits.argmax(1)[0] # (64, 64) predicted colours |
| ``` |
|
|
| Test-time training inside a game loop (`FinalPredictor` in the repository), the |
| agent, the Kaggle notebook, and the training commands are documented in the |
| [README](https://github.com/third-intelligence/WARC#quick-start) and in |
| [docs/reproducibility.md](https://github.com/third-intelligence/WARC/blob/main/docs/reproducibility.md). |
|
|
| ## Training data |
|
|
| The 340 human gameplay recordings of the 25 public ARC-AGI-3 games published by |
| the ARC Prize Foundation ([ARC-AGI-3 human dataset](https://arcprize.org/blog/arc-agi-3-human-dataset)). |
| The fold models use the recordings of their training games only; the all-25 |
| model uses one recording per game for validation and the rest for training. |
|
|
| ## Acknowledgements |
|
|
| WARC follows VARC ("ARC Is a Vision Problem!", Hu et al., 2025) and |
| reuses parts of the [VARC code](https://github.com/lillian039/VARC) (MIT). The |
| games and the recordings belong to the [ARC Prize Foundation](https://arcprize.org). |
| The repository lists every third-party component and its licence. |
|
|
| ## License |
|
|
| The weights and the code are released under |
| [CC0 1.0](https://creativecommons.org/publicdomain/zero/1.0/). Third-party |
| components keep their own licences. |
|
|
