Universal Reasoning Model

2025/12/16 by Zitian Gao, Gao, Zitian, Lynx Chen +13 · 8 voices
Computer Science · #Artificial Intelligence in Games #Explainable Artificial Intelligence (XAI) #Multimodal Machine Learning Applications

paper · pdf · doi:10.48550/arxiv.2512.14693

Abstract

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.

Citations

• Hierarchical Reasoning Model
• Scaling Exponents Across Parameterizations and Optimizers
• Grokked Transformers are Implicit Reasoners: A Mechanistic Journey to the Edge of Generalization
• Attention Is All You Need

Discussions

• Universal Reasoning Model (53.8% pass 1 ARC1 and 16.0% ARC 2) [hn, 131 points, 30 comments]
• Universal Reasoning Model (53.8% pass 1 ARC1 and 16.0% ARC 2) https://arxiv.org/abs/2512.14693 (https://news.ycombinator.com/item?id=46357458) [bsky, 0 points, 0 comments]
• https://bsky.app/profile/buzzing.cc.web.brid.gy/post/3manp3bhaqxx2 [bsky, 0 points, 0 comments]
• Universal Reasoning Model (53.8% pass 1 ARC1 and 16.0% ARC 2) https://arxiv.org/abs/2512.14693 https://news.ycombinator.com/item?id=46357458 [bsky, 0 points, 0 comments]
• Universal Reasoning Model (53.8% pass 1 ARC1 and 16.0% ARC 2) https://arxiv.org/abs/2512.14693 (https://news.ycombinator.com/item?id=46357458) [bsky, 0 points, 0 comments]
• ⚡ Hackernews Top story: Universal Reasoning Model (53.8% pass 1 ARC1 and 16.0% ARC 2) [bsky, 0 points, 0 comments]
• Universal Reasoning Model (53.8% pass 1 ARC1 and 16.0% ARC 2) https:// arxiv.org/abs/2512.14693 # arxiv [mastodon, 0 points, 0 comments]
• Universal Reasoning Model (53.8% pass 1 ARC1 and 16.0% ARC 2) https://arxiv.org/abs/2512.14693 [comments] [87 points] [bsky, 0 points, 0 comments]

Related

• Grokked Transformers are Implicit Reasoners: A Mechanistic Journey to the Edge of Generalization
• Can Transformers Really Do It All? On the Compatibility of Inductive Biases Across Tasks
• A Mechanistic Analysis of a Transformer Trained on a Symbolic Multi-Step Reasoning Task
• Sparse Universal Transformer
• On the Universality of Transformer Architectures; How Much Attention Is Enough?