A new paper from a Samsung AI researcher explains how a small network can beat massive Large Language Models (LLMs) in complex reasoning.
In the race for AI supremacy, the industry mantra has often been “bigger is better.” Tech giants have poured billions into creating ever-larger models, but according to Alexia Jolicoeur-Martineau of Samsung SAIL Montréal, a radically different and more efficient path forward is possible with the Tiny Recursive Model (TRM).
Using a model with just 7 million parameters, less than 0.01% of the size of leading LLMs, TRM achieves new state-of-the-art results on notoriously difficult benchmarks like the ARC-AGI intelligence test. Samsung’s work challenges the prevailing assumption that sheer scale is the only way to advance the capabilities of AI models, offering a more sustainable and parameter-efficient alternative.
Overcoming the limits of scale
While LLMs have shown incredible prowess in generating human-like text, their ability to perform complex, multi-step reasoning can be brittle. Because they generate answers token-by-token, a single mistake early in the process can derail the entire solution, leading to an invalid final answer.
Techniques like Chain-of-Thought, where a model “thinks out loud” to break down a problem, have been developed to mitigate this. However, these methods are computationally expensive, often require vast amounts of high-quality reasoning data that may not be available, and can still produce flawed logic. Even with these augmentations, LLMs struggle with certain puzzles where perfect logical execution is necessary.
Samsung’s work builds upon a recent AI model known as the Hierarchical Reasoning Model (HRM). HRM introduced a novel method using two small neural networks that recursively work on a problem at different frequencies to refine an answer. It showed great promise but was complicated, relying on uncertain biological arguments and complex fixed-point theorems that were not guaranteed to apply.
Instead of HRM’s two networks, TRM uses a single, tiny network that recursively improves both its internal “reasoning” and its proposed “answer”.
The model is given the question, an initial guess at the answer, and a latent reasoning feature. It first cycles through several steps to refine its latent reasoning based on all three inputs. Then, using this improved reasoning, it updates its prediction for the final answer. This entire process can be repeated up to 16 times, allowing the model to progressively correct its own mistakes in a highly parameter-efficient manner.
Counterintuitively, the research discovered that a tiny network with only two layers achieved far better generalisation than a four-layer version. This reduction in size appears to prevent the model from overfitting; a common problem when training on smaller, specialised datasets.
TRM also dispenses with the complex mathematical justifications used by its predecessor. The original HRM model required the assumption that its functions converged to a fixed point to justify its training method. TRM bypasses this entirely by simply back-propagating through its full recursion process. This change alone provided a massive boost in performance, improving accuracy on the Sudoku-Extreme benchmark from 56.5% to 87.4% in an ablation study.
Samsung’s model smashes AI benchmarks with fewer resources
The results speak for themselves. On the Sudoku-Extreme dataset, which uses only 1,000 training examples, TRM achieves an 87.4% test accuracy, a huge leap from HRM’s 55%. On Maze-Hard, a task involving finding long paths through 30×30 mazes, TRM scores 85.3% compared to HRM’s 74.5%.
Most notably, TRM makes huge strides on the Abstraction and Reasoning Corpus (ARC-AGI), a benchmark designed to measure true fluid intelligence in AI. With just 7M parameters, TRM achieves 44.6% accuracy on ARC-AGI-1 and 7.8% on ARC-AGI-2. This outperforms HRM, which used a 27M parameter model, and even surpasses many of the world’s largest LLMs. For comparison, Gemini 2.5 Pro scores only 4.9% on ARC-AGI-2.
The training process for TRM has also been made more efficient. An adaptive mechanism called ACT – which decides when the model has improved an answer enough and can move to a new data sample – was simplified to remove the need for a second, costly forward pass through the network during each training step. This change was made with no major difference in final generalisation.
This research from Samsung presents a compelling argument against the current trajectory of ever-expanding AI models. It shows that by designing architectures that can iteratively reason and self-correct, it is possible to solve extremely difficult problems with a tiny fraction of the computational resources.
See also: Google’s new AI agent rewrites code to automate vulnerability fixes
Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events including the Cyber Security Expo, click here for more information.
AI News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.
