Universal Reasoning Model

Source

• Raw Markdown: paper_universal-reasoning-model-2025.md
• PDF: paper_universal-reasoning-model-2025.pdf
• Preprint: arXiv 2512.14693
• Official code: UbiquantAI/URM

Core Claim

URM analyzes Universal Transformer variants for ARC-AGI and Sudoku, then adds short convolution and truncated backpropagation to improve small-model recursive reasoning.

Relevance To This Wiki

It is a direct modern descendant of UT for puzzle-like recursive reasoning, useful for separating recurrent inductive bias from elaborate architecture choices.

Limitations

The tasks are puzzle/reasoning tasks, not time-series state tracking. Reported pass@1 comparisons have a restricted setting that should be preserved when cited.

Foundation TSFM Relevance

Adjacent to latent-state refinement and dynamic compute; not direct TSFM evidence.

Links Into The Wiki

• Universal Reasoning Model
• Looped Transformers And Test-Time Memory
• Efficient Recurrent Sequence Models
• Time-Series Scaling And Efficiency
• Foundation Time-Series Model Research Agenda
• The Dragon Hatchling for an official but not open-reproduced BDH Sudoku narrative to compare with recursive Sudoku mechanisms.

Open Questions

• What matched-budget baseline should this source be compared against: unique-depth Transformer layers, recurrent state, explicit memory, or extra inference steps?
• Which claims transfer from token-sequence reasoning to multivariate time-series state tracking, event streams, or action-conditioned world models?
• Which gains come from UT-style recurrence, ConvSwiGLU nonlinearity, truncated backpropagation, or the ARC/Sudoku data protocol?