UC San Diego and Tsinghua University researchers develop breakthrough AI method that teaches small language models when to use tools versus internal knowledge, achieving 28% better accuracy while using fewer resources than larger models like GPT-4.