Samsung’s AI research team at the Advanced Institute of Technology in Montreal has developed a groundbreaking artificial intelligence model that defies the conventional wisdom that bigger models are inherently better. Their Tiny Recursive Model (TRM), with just 7 million parameters, has outshined much larger AI systems on some of the toughest reasoning benchmarks, challenging the dominance of giant language models from tech giants like Google and OpenAI.
The TRM achieves superior performance by employing a recursive reasoning approach. Instead of creating answers in a single pass, the model iteratively refines its responses by questioning and improving its own output multiple times. This process emulates human-like problem-solving, allowing the model to overcome errors early in reasoning and significantly enhance accuracy. The model smartly knows when to stop iterating, ensuring computational efficiency.
Despite its compact size a fraction of the scale and cost of leading models the TRM scored 44.6% accuracy on the ARC-AGI-1 benchmark, designed to evaluate fluid intelligence, beating Google’s Gemini 2.5 Pro, which scored 37%, and OpenAI’s o3-mini-high at 34.5%. It also achieved impressive results on concrete tasks such as solving Sudoku puzzles with 87.4% accuracy and maze navigation challenges at 85.3%.
Developed under the leadership of senior AI researcher Alexia Jolicoeur-Martineau, the TRM represents a shift from traditional large-scale model development. Instead of relying on billions or trillions of parameters, Samsung’s researchers demonstrated that smart architecture and recursive refining can produce better reasoning on less computational power, making advanced AI more accessible and efficient.
This breakthrough opens a new chapter in AI research focused on quality and innovation rather than sheer scale, providing a hopeful pathway for creating powerful, efficient, and sustainable AI systems.
This innovative model and its associated research paper have been made open source, encouraging wider exploration and adoption within the AI community.
