Recursive and latent-space reasoning models
Tiny networks that loop on a hidden state to solve hard puzzles from ~1,000 examples, an alternative to long chains of thought in huge models.
Reasoning depth can come from recursion in latent space rather than parameters or text tokens. HRM (27M parameters) and TRM (7M) beat much larger LLMs on Sudoku, mazes and ARC-AGI-1 with about 1,000 training examples, suggesting small specialised recursive models for hard verifiable puzzles.
Where it stands. Influential as a lens (refinement loops) but narrow. TRM components helped win the 2025 ARC Kaggle track, and none has been shown on open-ended language or agentic work (not found).
Evidence
For
- HRM (2025-06-26): two small networks recursing at different frequencies reach strong Sudoku, maze and ARC-AGI results at 27M parameters on ~1,000 examples.
- ARC Prize approximately reproduced HRM on its semi-private sets: 32% on ARC-AGI-1 and 2% on ARC-AGI-2 (2025-08-15).
- TRM (2025-10-06): a single 2-layer network with 7M parameters scores 45% on ARC-AGI-1 and 8% on ARC-AGI-2, above several far larger LLMs.
- TRM took first paper prize at ARC Prize 2025, and TRM components fed the top Kaggle entry (NVARC, 24% on ARC-AGI-2).
Against
- ARC Prize found HRM's hierarchical design had minimal impact versus a similar transformer; the outer refinement loop and memorizing evaluation-task solutions drove performance.
- Scale of the gap: TRM's 8% on ARC-AGI-2 compares with 54% for Gemini 3 Pro with a refinement harness (ARC Prize, Dec 2025).
- Not found: a recursive tiny model shown on open-ended language, coding or agentic tasks.
Milestones
- ARC Prize 2025 awards TRM first paper prize; CompressARC third ARC Prize Foundation · 5 December 2025
- TRM: Less is More, recursive reasoning with tiny networks Samsung SAIL Montreal · 6 October 2025
- ARC Prize analysis of HRM's hidden performance drivers ARC Prize Foundation · 15 August 2025
- HRM: Hierarchical Reasoning Model (arXiv v1) Sapient Intelligence · 26 June 2025
Who is working on it
- Guan Wang and co-authors (HRM), Sapient Intelligence
- Alexia Jolicoeur-Martineau (TRM), Samsung SAIL Montreal
The labs with the most milestones here are ARC Prize Foundation (2), Samsung SAIL Montreal (1) and Sapient Intelligence (1).
Sources
- arxiv.org/abs/2506.21734
- arxiv.org/abs/2510.04871
- arcprize.org/blog/hrm-analysis
- arcprize.org/blog/arc-prize-2025-results-analysis
This research bet was checked against its sources on 6 October 2026. How we check