A new research paper from Samsung AI challenges the conventional wisdom in the artificial intelligence industry that "bigger is better" for achieving advanced capabilities. Alexia Jolicoeur-Martineau of Samsung SAIL Montréal has introduced the Tiny Recursive Model (TRM), a radically different and highly efficient approach that uses a remarkably small network to outperform massive Large Language Models (LLMs) in complex reasoning tasks.

While tech giants invest billions into creating ever-larger models, TRM demonstrates that a model with just 7 million parameters—less than 0.01% of the size of leading LLMs—can achieve new state-of-the-art results on notoriously difficult benchmarks, including the ARC-AGI intelligence test. This work from Samsung directly questions the prevailing assumption that sheer scale is the sole path to advancing AI model capabilities, offering a more sustainable and parameter-efficient alternative.

The inherent limitations of current LLMs in complex reasoning tasks stem from their token-by-token generation process. A single error early in the sequence can invalidate an entire multi-step solution. Although techniques like Chain-of-Thought (CoT) have emerged to mitigate this by having models "think out loud," these methods are computationally expensive, often require extensive high-quality reasoning data, and can still lead to flawed logic. Even with these augmentations, LLMs frequently struggle with puzzles demanding perfect logical execution.

TRM’s development builds upon the foundation of a previous AI model known as the Hierarchical Reasoning Model (HRM). HRM introduced an innovative method where two small neural networks recursively refined a problem's solution at different frequencies. While promising, HRM was complex, relying on uncertain biological arguments and intricate fixed-point theorems whose applicability was not consistently guaranteed.

Distinguishing itself from HRM, TRM employs a single, tiny network that recursively refines both its internal "reasoning" and its proposed "answer." The model initiates its process by taking a question, an initial guess for the answer, and a latent reasoning feature. It then cycles through multiple steps to refine its latent reasoning based on these three inputs. Subsequently, this improved reasoning is used to update the prediction for the final answer. This entire iterative process can be repeated up to 16 times, enabling the model to progressively correct its own mistakes in a highly parameter-efficient manner.

Intriguingly, the research uncovered that a tiny network comprising only two layers exhibited superior generalization capabilities compared to a four-layer version. This reduction in size appears to effectively prevent the model from overfitting, a common challenge when training on smaller, specialized datasets. Furthermore, TRM streamlines the mathematical underpinnings of its predecessor by entirely dispensing with the complex justifications that HRM required regarding function convergence to a fixed point. Instead, TRM simply back-propagates through its complete recursion process, a simplification that alone delivered a significant boost in performance, improving accuracy on the Sudoku-Extreme benchmark from 56.5% to an impressive 87.4% in an ablation study.

The performance metrics of Samsung’s TRM are compelling. On the Sudoku-Extreme dataset, utilizing only 1,000 training examples, TRM achieved an 87.4% test accuracy, a substantial leap from HRM’s 55%. For Maze-Hard, a task requiring navigation through 30x30 mazes, TRM scored 85.3% compared to HRM’s 74.5%. Most notably, TRM made remarkable progress on the Abstraction and Reasoning Corpus (ARC-AGI), a benchmark specifically designed to assess fluid intelligence in AI. With merely 7 million parameters, TRM achieved 44.6% accuracy on ARC-AGI-1 and 7.8% on ARC-AGI-2. This not only surpasses HRM, which used a 27 million parameter model, but also outstrips many of the world's largest LLMs, including Gemini 2.5 Pro, which scored only 4.9% on ARC-AGI-2.

Training efficiency for TRM has also seen improvements. An adaptive mechanism known as Adaptive Computation Time (ACT), which determines when a model has sufficiently refined an answer before moving to a new data sample, was simplified. This modification eliminated the necessity for a second, resource-intensive forward pass through the network during each training step, without any significant compromise in final generalization performance.

In conclusion, this groundbreaking research from Samsung presents a powerful counter-argument to the current trend of perpetually expanding AI models. It decisively demonstrates that by architecting systems capable of iterative reasoning and self-correction, it is indeed possible to tackle extremely challenging problems with just a tiny fraction of the computational and parameter resources typically assumed necessary.