Samsung has just demonstrated how powerful specialized models can be, even at very small sizes. Despite being just a seven-million-parameter-sized AI model, which is a rounding error compared to massive foundational LLMs, it beats DeepSeek, o3-mini, and Gemini 2.5 Pro at sudoku, maze, and ARC-AGI puzzle solving.
Hard problems can be solved with very small models, as demonstrated by Alexia Jolicoeur-Martineau, a Senior AI Researcher at the Samsung SAIT AI Lab, in a new paper.
The researcher developed a novel recursive reasoning model approach to develop very small but very capable models for solving hard problems.
The Tiny Recursion Model (TRM) uses a very small neural model with just seven million parameters. This parameter count is thousands of times smaller compared to large foundational models, which contain hundreds of billions or sometimes even over a trillion parameters.
Despite the small size, a recursive model achieves “amazing scores.”
In the ARC-AGI-1 benchmark, which measures how effectively AI models solve specific puzzles, easy for humans but hard for AI systems, the TRM scored 45%. That's nearly the same as the GPT-5 (Low) score, and better compared to Gemini 2.5 Pro, Claude Opus 4.
On the ARC-AGI-2 benchmark, which further challenges systems to demonstrate both high adaptability and high efficiency, the TRM scored 8%, which was again ahead of major proprietary LLMs.
Where TRM truly shines is in the cost of the same task – it’s in a category of its own. To achieve the same score, the cost per task was just a fraction of a cent, whereas LLMs often required from a quarter to over a dollar to achieve the same result.
“The idea that one must rely on massive foundational models trained for millions of dollars by some big corporation in order to achieve success on hard tasks is a trap,” Jolicoeur-Martineau said.
The researcher suggests that there is too much focus on exploiting LLMs rather than devising and expanding new directions.
“With recursive reasoning, it turns out that ‘less is more’: you don’t always need to crank up model size in order for a model to reason and solve hard problems.”
How does the small TRM work?
As the researcher explained, the TRM simplifies recursive reasoning to its core essence – it recursively improves its predicted answer with a neural network.
When the TRM predicts the first answer, it tries to improve it for four more iterations, passing the previous question, answer, and reasoning as inputs repeatedly.
“This recursive process allows the model to progressively improve its answer (potentially addressing any errors from its previous answer) in an extremely parameter-efficient manner while minimizing overfitting,” Jolicoeur-Martineau explains.
According to the paper, increasing the number of layers of the model did not improve it. Instead, it decreased generalization due to overfitting, which occurs when the model learns the training data too closely and can’t make accurate predictions on unseen data.
The novel technique is simpler and significantly reduces the number of parameters required
Unlock more exclusive Cybernews content on YouTube.
Your email address will not be published. Required fields are markedmarked