@inproceedings{shu-etal-2026-loopvit,
title = "{L}oop{V}i{T}: Scaling Visual {ARC} with Looped Transformers",
author = "Shu, Wen-Jie and
Qiu, Xuerui and
Zhu, Rui-Jie and
Chen, Harold Haodong and
Liu, Yexin and
Yang, Harry",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2026",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-emnlp.1005/",
doi = "10.18653/2026.findings-emnlp.1005",
pages = "21271--21280",
ISBN = "979-8-89176-462-0",
abstract = "Recent ARC-AGI systems have shown that pure vision can solve abstract grid reasoning without language backbones, but most strong models still rely on fixed-depth feed-forward computation. We study whether recurrent computation is a better scaling axis for this setting. We propose Loop-ViT, a vision-native recurrent model that reuses a weight-tied Hybrid Block across multiple steps, with local convolutions for image-token refinement and attention for global rule induction. We also use an entropy-based exit rule to reduce unnecessary iterations at inference time. On ARC-AGI-1, Loop-ViT improves the parameter-efficiency trade-off over fixed-depth vision baselines and reaches 65.8{\%} accuracy with an 18M model. Overall, the results suggest that repeated computation is a useful inductive bias for ARC-style visual reasoning. The code is available at \href{https://github.com/WenjieShu/LoopViT}{https://github.com/WenjieShu/LoopViT}."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="shu-etal-2026-loopvit">
<titleInfo>
<title>LoopViT: Scaling Visual ARC with Looped Transformers</title>
</titleInfo>
<name type="personal">
<namePart type="given">Wen-Jie</namePart>
<namePart type="family">Shu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xuerui</namePart>
<namePart type="family">Qiu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Rui-Jie</namePart>
<namePart type="family">Zhu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Harold</namePart>
<namePart type="given">Haodong</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yexin</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Harry</namePart>
<namePart type="family">Yang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: EMNLP 2026</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hung-Yi</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sunipa</namePart>
<namePart type="family">Dev</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jessy</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Desmond</namePart>
<namePart type="family">Elliott</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-462-0</identifier>
</relatedItem>
<abstract>Recent ARC-AGI systems have shown that pure vision can solve abstract grid reasoning without language backbones, but most strong models still rely on fixed-depth feed-forward computation. We study whether recurrent computation is a better scaling axis for this setting. We propose Loop-ViT, a vision-native recurrent model that reuses a weight-tied Hybrid Block across multiple steps, with local convolutions for image-token refinement and attention for global rule induction. We also use an entropy-based exit rule to reduce unnecessary iterations at inference time. On ARC-AGI-1, Loop-ViT improves the parameter-efficiency trade-off over fixed-depth vision baselines and reaches 65.8% accuracy with an 18M model. Overall, the results suggest that repeated computation is a useful inductive bias for ARC-style visual reasoning. The code is available at \hrefhttps://github.com/WenjieShu/LoopViThttps://github.com/WenjieShu/LoopViT.</abstract>
<identifier type="citekey">shu-etal-2026-loopvit</identifier>
<identifier type="doi">10.18653/2026.findings-emnlp.1005</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.1005/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>21271</start>
<end>21280</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T LoopViT: Scaling Visual ARC with Looped Transformers
%A Shu, Wen-Jie
%A Qiu, Xuerui
%A Zhu, Rui-Jie
%A Chen, Harold Haodong
%A Liu, Yexin
%A Yang, Harry
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Findings of the Association for Computational Linguistics: EMNLP 2026
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-462-0
%F shu-etal-2026-loopvit
%X Recent ARC-AGI systems have shown that pure vision can solve abstract grid reasoning without language backbones, but most strong models still rely on fixed-depth feed-forward computation. We study whether recurrent computation is a better scaling axis for this setting. We propose Loop-ViT, a vision-native recurrent model that reuses a weight-tied Hybrid Block across multiple steps, with local convolutions for image-token refinement and attention for global rule induction. We also use an entropy-based exit rule to reduce unnecessary iterations at inference time. On ARC-AGI-1, Loop-ViT improves the parameter-efficiency trade-off over fixed-depth vision baselines and reaches 65.8% accuracy with an 18M model. Overall, the results suggest that repeated computation is a useful inductive bias for ARC-style visual reasoning. The code is available at \hrefhttps://github.com/WenjieShu/LoopViThttps://github.com/WenjieShu/LoopViT.
%R 10.18653/2026.findings-emnlp.1005
%U https://aclanthology.org/2026.findings-emnlp.1005/
%U https://doi.org/10.18653/2026.findings-emnlp.1005
%P 21271-21280
Markdown (Informal)
[LoopViT: Scaling Visual ARC with Looped Transformers](https://aclanthology.org/2026.findings-emnlp.1005/) (Shu et al., Findings 2026)
ACL
- Wen-Jie Shu, Xuerui Qiu, Rui-Jie Zhu, Harold Haodong Chen, Yexin Liu, and Harry Yang. 2026. LoopViT: Scaling Visual ARC with Looped Transformers. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 21271–21280, Budapest, Hungary. Association for Computational Linguistics.