Less is More: Recursive Reasoning with Tiny Networks

Source

• Raw Markdown: paper_tiny-recursive-model-2025.md
• PDF: paper_tiny-recursive-model-2025.pdf
• Preprint: arXiv 2510.04871

Core Claim

TRM simplifies HRM into a single tiny recursive network and reports stronger generalization on Sudoku, Maze, and ARC-AGI style tasks with about 7M parameters.

Relevance To This Wiki

TRM is the minimalist recursive-reasoning counterpoint to HRM: it asks how much of the gain comes from recurrence and deep supervision rather than biological hierarchy.

Limitations

It is small-model puzzle evidence. Its lesson should be translated as recurrence and supervision structure, not as a general replacement for sequence-model scale.

Foundation TSFM Relevance

Important for the fixed-FLOPs and small-recursive-model thread, especially when comparing hierarchy versus repeated state refinement.

Links Into The Wiki

• Tiny Recursive Model
• Looped Transformers And Test-Time Memory
• Efficient Recurrent Sequence Models
• Time-Series Scaling And Efficiency
• Foundation Time-Series Model Research Agenda

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?
• Does the answer-refinement loop remain strong outside small discrete puzzle states with full observability?