A research breakthrough from Samsung’s SAIL Montréal is challenging the core assumption that more powerful AI requires more parameters. Led by Alexia Jolicoeur-Martineau, the team has released the Tiny Recursive Model (TRM), an open-source model that dramatically outperforms both massive LLMs and the previously state-of-the-art Hierarchical Reasoning Model (HRM) on hard puzzle and reasoning benchmarks, all with dramatically fewer parameters and resources.
The “Secret Sauce”: Recursive Refinement
Instead of scaling up, TRM gets smarter by “thinking harder” through a process of recursion. Here’s how it works:
- A Single, Tiny Network: TRM uses just one compact neural network (as small as two layers) and loops through it repeatedly.
- Self-Correcting Thoughts: With each loop, it dynamically updates both its internal “reasoning state” and its final answer, allowing it to self-correct and refine its solution.
- Knows When to Stop: A learned “adaptive halting” mechanism tells the model when it has reached a good answer, saving computation.
This approach is both simpler and more efficient than its predecessor, the Hierarchical Reasoning Model (HRM), which relied on a more complex two-network setup.
Stunning Performance: Small Model, Massive Gains
The benchmark results make TRM’s capabilities clear. It excels precisely where giant LLMs struggle: abstract, logical, and spatial reasoning tasks.
Key Results Explained:
- Sudoku & Mazes: On notoriously difficult 9×9 Sudoku puzzles, TRM achieved 87.4% accuracy, dwarfing HRM’s 55% and leaving LLMs in the dust. It showed similar dominance on complex 30×30 mazes.
- ARC-AGI Benchmarks: These tests for abstract reasoning are considered proxies for AGI-like intelligence. Here, the 7-million-parameter TRM outperformed models with over 95,000 times its parameters, including DeepSeek R1 and Gemini 2.5 Pro.
Why This Is a Landmark Achievement
- Turns “Bigger is Better” on Its Head: TRM proves that architectural ingenuity can be a more powerful lever than brute-force scaling, especially for logical reasoning.
- Unlocks Efficient Generalisation: Its ability to learn from very few examples (only 1,000 training Sudoku puzzles) suggests a path toward AI that generalizes more like humans.
- Open and Accessible: As an open-source project, TRM provides a blueprint for the entire research community to build upon, potentially accelerating the development of efficient AI.
Expert Perspective
Experts highlight that TRM’s “latent reasoning” is more robust than the brittle, language-dependent “chain-of-thought” used by LLMs. It demonstrates a form of non-verbal, deep thought process that is critical for true reasoning.
Conclusion
The Tiny Recursive Model is more than just a state-of-the-art model. It offers compelling evidence that the next great leaps in artificial intelligence may come not from building ever-larger models, but from designing more clever, efficient, and brain-inspired architectures.
Reference:
Jolicoeur-Martineau, A., Groleau, A., Ma, K., Zhang, M. and Liu, Z. (2024). Tiny Recursive Model for Systematic Generalization. arXiv preprint. Available at: https://arxiv.org/abs/2510.04871 [Accessed 10 October 2025]
