Samsung's Compact AI Model Outperforms Larger Rivals in Reasoning

A new paper by a Samsung AI researcher outlines how a compact network can outperform massive large language models (LLMs) in tackling complex reasoning tasks.

In the race for AI dominance, the prevailing industry mantra has been "bigger is better." While tech giants invest billions in developing increasingly large models, Alexia Jolicoeur-Martineau of Samsung SAIL Montréal proposes a different, more efficient approach using the Tiny Recursive Model (TRM).

With just 7 million parameters—less than 0.01% the size of top LLMs—TRM achieves new state-of-the-art results on notoriously challenging benchmarks like the ARC-AGI intelligence test. Samsung’s findings challenge the common belief that scale is the only path to advancing AI, presenting a more sustainable and parameter-efficient alternative.

Overcoming the limits of scale

Although LLMs excel at generating human-like text, their ability to handle complex, multi-step reasoning is often fragile. Since they generate responses token by token, one error early on can compromise the entire solution, resulting in an incorrect final answer.

Techniques such as Chain-of-Thought, in which a model breaks down a problem step by step, aim to mitigate this. However, these approaches are computationally costly, often require substantial high-quality reasoning data, and can still produce flawed logic. Even with these enhancements, LLMs have difficulty solving puzzles that demand flawless logical execution.

Samsung’s research builds on the recent Hierarchical Reasoning Model (HRM). HRM used two small neural networks that recursively refine answers at different frequencies. Although promising, the model was complex, relying on uncertain biological arguments and fixed-point theorems that were not always applicable.

Instead of HRM’s dual-network structure, TRM employs a single, compact network that recursively enhances both its internal reasoning and its proposed answer.

The model receives the question, an initial answer guess, and a latent reasoning feature. It then cycles through multiple steps to refine its reasoning based on all three inputs. Using this improved reasoning, it updates its final answer prediction. This entire process can repeat up to 16 times, enabling the model to progressively self-correct in a highly parameter-efficient way.

Counterintuitively, the study found that a two-layer network generalized much better than a four-layer version. The smaller design appears to prevent overfitting—a common issue when training on limited, specialized datasets.

TRM also eliminates the complex mathematical assumptions of its predecessor. The original HRM model had to assume function convergence to a fixed point to justify its training. TRM bypasses this by back-propagating through its full recursion process—a change that vastly improved performance, raising Sudoku-Extreme benchmark accuracy from 56.5% to 87.4% in ablation tests.

Samsung’s model smashes AI benchmarks with fewer resources

The results are striking. On the Sudoku-Extreme dataset, which uses only 1,000 training examples, TRM achieves 87.4% test accuracy—a major leap from HRM’s 55%. On Maze-Hard, which involves navigating long paths through 30×30 mazes, TRM scores 85.3%, compared to HRM’s 74.5%.

Most notably, TRM makes significant progress on the Abstraction and Reasoning Corpus (ARC-AGI), a benchmark designed to evaluate true fluid intelligence in AI. With just 7M parameters, TRM reaches 44.6% accuracy on ARC-AGI-1 and 7.8% on ARC-AGI-2. That beats HRM, which used 27M parameters, and even surpasses many of the world’s largest LLMs. For comparison, Gemini 2.5 Pro scores only 4.9% on ARC-AGI-2.

TRM’s training process has also been optimized. An adaptive mechanism called ACT—which determines when the model has sufficiently improved an answer to move on—was simplified, eliminating the need for a costly second forward pass during each training step. This adjustment did not harm overall generalization.

Samsung’s research provides a strong counterpoint to the trend of building ever-larger AI models. It demonstrates that by designing architectures capable of iterative reasoning and self-correction, extremely difficult problems can be solved using only a tiny fraction of the computational resources.

See also: Google’s new AI agent rewrites code to automate vulnerability fixes

Interested in learning more about AI and big data from industry leaders? Attend the AI & Big Data Expo in Amsterdam, California, and London. This comprehensive event is part of TechEx and runs alongside other major tech events like the Cyber Security Expo. Click here for details.

AI News is powered by TechForge Media. Discover other upcoming enterprise technology events and webinars here.