In the fast-paced world of artificial intelligence, the pursuit of models capable of tackling complex reasoning tasks with both efficiency and accuracy is a monumental challenge. Enter the Hierarchical Reasoning Model (HRM) and its impressive successor, the Tiny Recursive Model (TRM). These models are making significant waves, particularly in their ability to solve intricate puzzles that often stump even the largest language models (LLMs).

Hierarchical Reasoning Model (HRM): A Biologically Inspired Start

The journey began with the Hierarchical Reasoning Model (HRM), a novel approach to complex reasoning inspired by the sophisticated networks of human and animal brains. HRM ingeniously employed two small neural networks, recurring at different frequencies, to perform its computations. Its core innovation was deep supervision, a technique that allowed the model to iteratively refine its answers. By reusing past computations, HRM could adjust and self-correct its reasoning process over multiple steps, much like a person reviewing their work.

This recursive methodology enabled HRM to mimic the reasoning depth typically found in much larger networks, but crucially, without the heavy computational burden of backpropagating through countless layers. HRM showcased remarkable performance on challenging puzzle tasks such as Sudoku, Maze navigation, and ARC-AGI. It often outshone LLMs, despite being trained with a mere 27 million parameters and approximately 1000 examples—a tiny footprint by comparison.

However, independent analyses later revealed a pivotal insight: the true power of HRM didn't lie in its complex hierarchical structure or elaborate recursion, but predominantly in the iterative improvement facilitated by deep supervision. This understanding paved the way for a simpler, yet even more effective, successor.

Tiny Recursive Model (TRM): Less is More, Smarter is Better

Building on the foundational insights gleaned from HRM, researchers at Samsung AI Lab developed the Tiny Recursive Model (TRM). TRM embodies the philosophy of "less is more," presenting a profoundly effective, yet much simpler, recursive reasoning approach. Unlike HRM's dual-network, hierarchical design, TRM employs a single, tiny two-layer network. This radical simplification has led to significantly higher generalization across a diverse range of problems.

TRM's key innovations and advantages are nothing short of astounding:

• Exceptional Efficiency: TRM operates with an astonishingly small number of parameters—as little as 7 million. This represents less than 0.01% of the parameters found in many large language models, making it incredibly computationally efficient and accessible.
• Superior Performance: Despite its diminutive size, TRM has demonstrated superior performance on complex reasoning tasks, including Sudoku and ARC-AGI puzzles. It consistently outperforms both HRM and many large LLMs such as Deepseek R1, o3-mini, and even Gemini 2.5 Pro in terms of test accuracy (e.g., 45% on ARC-AGI-1 and 8% on ARC-AGI-2).
• Simplified Design: TRM sheds the complexities of HRM. It requires no fixed-point theorem, intricate biological justifications, or a hierarchical structure. Instead, it distills the effective components of iterative refinement into a streamlined, elegant architecture.
• Enhanced Generalization: By focusing on the core principle of iterative refinement with a simplified network, TRM achieves better generalization capabilities, enabling it to solve a broader spectrum of complex problems with greater adaptability.

Conclusion

The evolution from HRM to TRM marks a critical juncture in the pursuit of intelligent reasoning systems. While HRM pioneered the power of iterative reasoning and deep supervision, TRM unequivocally demonstrates that by simplifying the architecture and focusing on the most effective mechanisms, a "tiny" model can not only match but significantly surpass the capabilities of much larger and more complex counterparts. The Tiny Recursive Model represents a promising direction for developing highly efficient and effective AI systems, capable of tackling complex reasoning tasks with unprecedented resource efficiency. This breakthrough could unlock advanced reasoning capabilities in resource-constrained environments and pave the way for more interpretable, robust, and sustainable AI solutions.