IAAILoop-ViT
LoopViT: Scaling Visual ARC with Looped Transformers
A recursive vision architecture for ARC-AGI that iterates a weight-tied hybrid block of local convolutions and global attention as a latent chain of thought, with a parameter-free dynamic exit when predictive entropy shows the state has settled.
Inside the method
Grid → [ Weight-tied conv + attention block ] × r (entropy exit) → output grid
Simplified conceptual schematic. Consult the paper for the complete architecture.
- Recurrence family
- Adaptive depth
- Depth control
- Parameter-free entropy-based dynamic exit
- KV / state strategy
- Not applicable (vision model)
Reading note
An 18M model reaches 65.8% on ARC-AGI-1, outperforming 73M-parameter ensembles.
Sources checked 2026-10-28. This catalog does not imply independent reproduction.
Related works
Cite this work
@misc{loopvit2026,
title = {LoopViT: Scaling Visual ARC with Looped Transformers},
author = {Wen-Jie Shu and Xuerui Qiu and Rui-Jie Zhu and Harold Haodong Chen and Yexin Liu and Harry Yang},
year = {2026},
eprint = {2602.02156},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2602.02156}
}