Researchers Found a Better Way to Teach Large Language Models New Skills