A recent study from Samsung Advanced Institute of Technology’s AI Lab introduces a groundbreaking small AI model called the Tiny Recursive Model (TRM). With only 7 million parameters, TRM achieved notable performance on the ARC-AGI benchmarks, scoring 45% accuracy on ARC-AGI 1, surpassing larger models like Google’s Gemini 2.5 Pro (37%) and OpenAI’s o3-mini-high (34.5%). In the challenging ARC-AGI-2 benchmark, TRM attained 7.8% accuracy, outpacing Gemini and o3-mini-high. Developed efficiently using just four NVIDIA H-100 GPUs for under $500, TRM showcases a recursive approach to improve answers iteratively, minimizing overfitting. This evidence supports the thesis “Less is More,” suggesting that smaller models can outperform larger ones in specific tasks. Industry experts, including Deedy Das of Menlo Ventures, endorse this shift towards specific, cost-effective models for enhanced automation solutions. For further insights and details, refer to the study linked in the report.
Source link
