Hierarchical Reasoning Model

Source

• Raw Markdown: paper_hierarchical-reasoning-model-2025.md
• PDF: paper_hierarchical-reasoning-model-2025.pdf
• Preprint: arXiv 2506.21734
• Official code: sapientinc/HRM

Core Claim

HRM uses interdependent high-level and low-level recurrent modules operating at different timescales to solve reasoning tasks in a single forward pass without explicit chain-of-thought supervision.

Relevance To This Wiki

HRM is a landmark recursive-reasoning branch adjacent to UT and looped Transformers: it shows that tiny recurrent systems can solve hard puzzle tasks without large language-model pretraining.

Limitations

The impressive evidence is concentrated in puzzle and ARC-style domains. Transfer to temporal world models requires a separate observation, action, and state contract.

Foundation TSFM Relevance

Important for hierarchical latent-state updates and fast/slow processing, but still architecture background rather than direct multivariate time-series evidence.

Links Into The Wiki

• Hierarchical Reasoning 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?
• Which part of the result is hierarchy, deep supervision, recurrence, or task-specific puzzle augmentation?