Hierarchical Reasoning Model (HRM)
A 27M-parameter two-timescale recurrent model, trained on about 1,000 examples, solves Sudoku and mazes and scores on ARC without chain of thought.
Two coupled recurrent modules (slow planner, fast worker) iterate to depth without token-level CoT. Viral for ARC-AGI results at tiny scale; the ARC Prize team's follow-up analysis questioned how much came from the hierarchy versus data augmentation and refinement loops.
- Date
- Thursday, 26 June 2025
- Lab
- Sapient Intelligence (per authors' project)
- Kind
- paper
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| Parameters | 27M ~1,000 training examples; ARC, Sudoku, mazes | authors |
Lead author Guan Wang; arXiv v1 2025-06-26. Follow-up analyses questioned how much of the ARC result came from the hierarchy versus data augmentation; that critique was not verified here. Superseded in simplicity by TRM.
Sources
This record was checked against its sources on 6 October 2026. How we check