Researchers developed practical training recipes for looped language models, reducing training budgets and improving performance on various benchmarks. They also introduced a method to convert pretrained dense models into looped ones, achieving gains on multiple tasks.