Overview
- Samsung SAIL Montreal published the Tiny Recursive Model with an arXiv paper and released full code and training details on GitHub under the MIT license.
- TRM iteratively feeds its outputs back as inputs and uses deep supervision plus adaptive halting to improve predictions over multiple passes.
- Reported scores include 87.4% on Sudoku-Extreme, 85% on Maze-Hard, 45% on ARC-AGI-1, and 8% on ARC-AGI-2.
- Coverage and the authors state the model matches or surpasses much larger LLMs such as Google’s Gemini 2.5 Pro, OpenAI’s o3-mini, and DeepSeek R1 on these structured puzzles.
- At 7 million parameters, the system can run on commodity hardware with lower energy use, though it has not been evaluated on open-ended language or perceptual tasks.