Investigating the Robustness of LLMs on Math Word Problems