“Larger models benefit more from thinking time provided by Chain-of-Thought prompting when solving math problems.”