IMU-1: Sample-Efficient Pre-training of Small Language Models Paper โข 2602.02522 โข Published Jan 25 โข 8