The Diminishing Returns of Masked Language Models to Science