Samsung AI researchers have made a groundbreaking leap in artificial intelligence, challenging the long-held industry belief that bigger models yield better results. Their Tiny Recursive Model (TRM), with just 7 million parameters, has outperformed massive language models thousands of times its size on advanced reasoning tasks—proving that intelligent design can outshine raw computational scale.
Led by Alexia Jolicoeur-Martineau at Samsung SAIL Montreal, the research is detailed in the paper “Less is More: Recursive Reasoning with Tiny Networks.” Unlike tech giants investing billions into models with hundreds of billions of parameters, Samsung’s TRM delivers superior performance on complex benchmarks using less than 0.01% of the computational resources.
🧠 Outsmarting Giants on AI’s Toughest Challenges
TRM’s results have stunned the AI community. On the ARC-AGI-1 benchmark, which tests fluid intelligence, the model scored 44.6%—beating much larger models like DeepSeek-R1, Google’s Gemini 2.5 Pro, and OpenAI’s o3-mini. On the more demanding ARC-AGI-2 test, TRM achieved 7.8%, surpassing Gemini 2.5 Pro’s 4.9%.
Its capabilities extend to practical problem-solving as well. TRM reached 87.4% accuracy on Sudoku-Extreme puzzles after training on just 1,000 examples, and scored 85.3% on maze navigation tasks involving 30×30 grids—showcasing impressive generalization.
🔁 Recursive Thinking: The Tiny Model’s Big Advantage
TRM’s strength lies in its recursive reasoning strategy, which mimics human problem-solving more closely than traditional AI models. Instead of producing answers in a single pass, TRM iteratively refines its solutions using an internal “scratchpad,” revisiting and improving its reasoning up to 16 times.
This method tackles a major flaw in conventional models—the tendency for early errors to snowball through the solution. “Relying on massive foundational models built at enormous cost isn’t the only path to solving hard problems,” Jolicoeur-Martineau noted. The findings suggest that recursive logic may be the key to mastering abstract reasoning where even top-tier generative models falter.
TRM also simplifies its predecessor, the Hierarchical Reasoning Model, which relied on dual networks and complex math. Samsung’s new approach uses a streamlined two-layer network that recursively enhances both its thought process and final answers.
Leave a Reply