HRM and TRM, two small recursive reasoners with 27M and 7M parameters
The Hierarchical Reasoning Model (HRM; arXiv v1 2025-06-26, revised through v4 on 2026-09-30) is a 27M-parameter recurrent architecture with two interdependent modules, one for slow abstract planning…
- Date
- 26 June 2025
- Who
- HRM by Guan Wang et al. (nine authors); TRM by Alexia Jolicoeur-Martineau (sole author)
- Confidence
- High on what the papers claim; Low on generality
- Deep dive
- Reasoning II, from o1 and o3 to DeepSeek-R1 and the labs that replicated them
Tier: Supporting · Significance: 3/5 · Org(s): HRM by Guan Wang et al. (nine authors); TRM by Alexia Jolicoeur-Martineau (sole author) · Confidence: High on what the papers claim; Low on generality The Hierarchical Reasoning Model (HRM; arXiv v1 2025-06-26, revised through v4 on 2026-09-30) is a 27M-parameter recurrent architecture with two interdependent modules, one for slow abstract planning and one for fast detailed computation, that solves sequential reasoning tasks in a single forward pass from about 1,000 training samples, with no pretraining or chain-of-thought data; the authors report near-perfect Sudoku and maze results and strong results on ARC (arXiv 2506.21734). The Tiny Recursive Model (TRM; arXiv v1 2025-10-06) replaces it with a single 2-layer network of 7M parameters and reports 45% on ARC-AGI-1 and 8% on ARC-AGI-2, which its abstract says beats most LLMs, including DeepSeek R1, o3-mini and Gemini 2.5 Pro, with under 0.01% of their parameters (arXiv 2510.04871). The results come with caveats. These are narrow, task-trained models evaluated on puzzle sets and are not general reasoners, and TRM's 8% on ARC-AGI-2 sits far below the 2026 frontier results (B08-43b). They matter as an alternative to token-level chain of thought, namely recursion in latent space, beside the looped-transformer work of B08-17b in the legibility debate that reaches B08-50. Sources: arXiv 2506.21734 · arXiv 2510.04871