mlsanityData to 14 Oct 2026

Universal Reasoning Model

arXiv 2512.14693 ↗First posted 16 Dec 2025cs.AIcited by 11

On the map: LLM reasoning → Reasoning benchmarks and evaluation

Abstract (arXiv)

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.

Key facts. Cited by 11 papers in the map (data through 14 Oct 2026). Placed in Reasoning benchmarks and evaluation, one of 17 directions in LLM reasoning. Builds on 13 of its 23 references that are in the map. First posted on arXiv 16 Dec 2025.

Builds on

13 of 23 references.

Closest ancestors · 2

Same direction, newest first

  1. 2025Hierarchical Reasoning Models: Perspectives and Misconceptions30 Sep
  2. 2024Grokked Transformers are Implicit Reasoners: A Mechanistic Journey to the Edge of Generalization23 May

Other references · 11

  1. 2020GLU Variants Improve Transformercited by 1,804
  2. 2018Universal Transformerscited by 647
  3. 2016Adaptive Computation Time for Recurrent Neural Networkscited by 506
  4. 2025Muon is Scalable for LLM Trainingcited by 428
  5. 2025ARC-AGI-2: A New Challenge for Frontier AI Reasoning Systemscited by 190
  6. 2025Hierarchical Reasoning Modelcited by 177
  7. 2024ARC Prize 2024: Technical Reportcited by 104
  8. 2025Less is More: Recursive Reasoning with Tiny Networkscited by 45
  9. 2017Unbiasing Truncated Backpropagation Through Timecited by 37
  10. 2025ARC Is a Vision Problem!cited by 24
  11. 2025What Makes Looped Transformers Perform Better Than Non-Recursive Onescited by 3

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2026: 11.

11 papers.

  1. 2026Fixed-Point Reasoners: Stable and Adaptive Deep Looped Transformerscited by 28
  2. 2026Step-resolved data attribution for looped transformerscited by 5
  3. 2026Universal Transformers Need Memory: Depth-State Trade-offs in Adaptive Recursive Reasoningcited by 5
  4. 2026One Step Forward and K Steps Back: Better Reasoning with Denoising Recursion Modelscited by 3
  5. 2026Recursive Inference Machines for Neural Reasoningcited by 2
  6. 2026Steering Recurrent Reasoners at Inference Time with Readout Feedbackcited by 1
  7. 2026LoopCD: Loop-wise Contrastive Decoding for Improving Reasoning in Looped Language Modelscited by 0
  8. 2026Structured Sparse Memory for Recurrent Reasoningcited by 0
  9. 2026What Limits Recursive Reasoning Models: Optimization, Architecture and Test-Time Scalingcited by 0
  10. 2026Unrolled Flow Models for Reasoningcited by 0
  11. 2026Scaling to Tens of Thousands of Test-Time Iterations with Loop-Native Attention Residualscited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2025Liquid Reasoning Transformers: A Sudoku-Based Prototype for Chess-Scale Algorithmic Tasks
  2. 2023Sparse Universal Transformer
  3. 2024Reasoning in Transformers -- Mitigating Spurious Correlations and Reasoning Shortcuts
  4. 2026Transformers Linearly Represent Highly Structured World Models
  5. 2026Modality-Driven Search with Holistic Trace Judging for ARC-AGI-2
  6. 2025Emergence of Minimal Circuits for Indirect Object Identification in Attention-Only Transformers
  7. 2024Seq-VCR: Preventing Collapse in Intermediate Transformer Representations for Enhanced Reasoning
  8. 2024UTMath: Math Evaluation with Unit Test via Reasoning-to-Coding Thoughts