Tiny Recursive Model (TRM)
A 7M-parameter, 2-layer recursive network reaches 45% on ARC-AGI-1 and 8% on ARC-AGI-2, above several far larger LLMs.
Strips HRM down to one tiny network that recursively refines its own answer. Gets higher generalisation than HRM on Sudoku, mazes and ARC. Claims score above DeepSeek R1, o3-mini and Gemini 2.5 Pro on these puzzles with under 0.01% of the parameters.
- Date
- Monday, 6 October 2025
- Lab
- Samsung SAIL Montreal
- Kind
- paper
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| ARC-AGI-1 / ARC-AGI-2 test accuracy | 45% / 8% 7M parameters; author-reported | authors |
Single author Alexia Jolicoeur-Martineau. Task-specific training on ARC-style data, so not comparable to general LLM scores.
Sources
This record was checked against its sources on 6 October 2026. How we check