NCAs solve Sudoku straight from raw pixels, reasoning end-to-end on a 256x256 canvas
BertChakovsky · x · 2026-10-02
A new Neural Cellular Automata demo, "Visual Sudoku", feeds the model a raw-pixel image of an unsolved Sudoku and has it produce a valid solved board end-to-end on a 256x256 canvas.
The task requires simultaneous classification, reasoning, and generation over pixels with no symbolic intermediate representation — showing that NCA-style locally-interacting models can carry out multi-step reasoning directly in pixel space. An interactive demo accompanies the release.
More from Research
- AMap open-sources ABot-Recon: streaming 3D reconstruction from video with a 12-frame local context — rsasaki0109 · 2026-10-02
- Neuralink pretrains decoders on 50,000+ hours of neural data—dataset may be the real moat — CurieuxExplorer · 2026-10-02
- CyberGym cybersecurity benchmark effectively saturated on verified task subset — aryaman2020 · 2026-10-02
- SemEval-2027 Task 9 calls for teams on 4-language multimodal news framing analysis — preslav_nakov · 2026-10-02
- Alternating prompt and model upgrades lift science agent from 42% to 73% — rohanpaul_ai · 2026-10-02
- ISMIR paper teaches a transformer to play in 12 jazz piano legends' styles — umpedronosapato · 2026-10-02