In the competitive AI landscape, major tech firms are vying to create better AI models. Notably, Samsung’s Senior AI Researcher, Alexia Jolicoeur-Martineau, has developed the Tiny Recursion Model (TRM), a breakthrough neural network with only 7 million parameters. Despite its size, TRM has outperformed larger LLMs (Large Language Models) like OpenAI’s o3-mini and Google’s Gemini 2.5 Pro in reasoning benchmarks, achieving scores of 45% on ARC-AGI-1 and 8% on ARC-AGI-2. The model’s code is accessible on GitHub under the MIT License, allowing for commercial use. Jolicoeur-Martineau argues against the reliance on massive foundational models, advocating instead for innovative approaches that optimize efficiency. TRM employs a streamlined version of the Hierarchical Reasoning Model (HRM) to effectively handle structured problems, simplifying the process into a two-layer framework. This model highlights that a smaller system can deliver significant results without exorbitant costs, thereby revolutionizing AI development.
Source link
Share
Read more