The new technique lets LLMs adapt computation to problem difficulty, reducing energy use and enabling smaller models to ...