Format Matters: The Robustness of Multimodal LLMs in Reviewing Evidence from Tables and Charts