Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Gao, Zitian, Chen, Lynx, Xiao, Yihao, Xing, He, Tao, Ran, Luo, Haoming, Zhou, Joey, Dai, Bryan
Format: Preprint
Veröffentlicht: 2025
Schlagworte:
Online-Zugang:https://arxiv.org/abs/2512.14693
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866908732496019456
author Gao, Zitian
Chen, Lynx
Xiao, Yihao
Xing, He
Tao, Ran
Luo, Haoming
Zhou, Joey
Dai, Bryan
author_facet Gao, Zitian
Chen, Lynx
Xiao, Yihao
Xing, He
Tao, Ran
Luo, Haoming
Zhou, Joey
Dai, Bryan
contents Universal transformers (UTs) have been widely used for complex reasoning tasks such as ARC-AGI and Sudoku, yet the specific sources of their performance gains remain underexplored. In this work, we systematically analyze UTs variants and show that improvements on ARC-AGI primarily arise from the recurrent inductive bias and strong nonlinear components of Transformer, rather than from elaborate architectural designs. Motivated by this finding, we propose the Universal Reasoning Model (URM), which enhances the UT with short convolution and truncated backpropagation. Our approach substantially improves reasoning performance, achieving state-of-the-art 53.8% pass@1 on ARC-AGI 1 and 16.0% pass@1 on ARC-AGI 2. Our code is avaliable at https://github.com/UbiquantAI/URM.
format Preprint
id arxiv_https___arxiv_org_abs_2512_14693
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Universal Reasoning Model
Gao, Zitian
Chen, Lynx
Xiao, Yihao
Xing, He
Tao, Ran
Luo, Haoming
Zhou, Joey
Dai, Bryan
Artificial Intelligence
Universal transformers (UTs) have been widely used for complex reasoning tasks such as ARC-AGI and Sudoku, yet the specific sources of their performance gains remain underexplored. In this work, we systematically analyze UTs variants and show that improvements on ARC-AGI primarily arise from the recurrent inductive bias and strong nonlinear components of Transformer, rather than from elaborate architectural designs. Motivated by this finding, we propose the Universal Reasoning Model (URM), which enhances the UT with short convolution and truncated backpropagation. Our approach substantially improves reasoning performance, achieving state-of-the-art 53.8% pass@1 on ARC-AGI 1 and 16.0% pass@1 on ARC-AGI 2. Our code is avaliable at https://github.com/UbiquantAI/URM.
title Universal Reasoning Model
topic Artificial Intelligence
url https://arxiv.org/abs/2512.14693