LoopViT: Scaling Visual ARC with Looped Transformers
Abstract
We introduce Loop-ViT, a weight-tied (looped) transformer for Visual ARC that scales reasoning via iterative inference. By trading model width for “thinking time”, Loop-ViT reaches a new Pareto frontier on ARC-AGI, achieving strong accuracy with significantly fewer parameters.
Type
Publication
EMNLP 2026 Findings
Accepted at EMNLP 2026 Findings.
Figure: LoopViT overview.
