Beyond Via: Large-Scale Analysis Detects LLM-Driven Word Frequency Shifts in arXiv Papers — 'Beyond' and 'Via' Up, 'The' and 'Of' Down

arXiv·medium signal

Geng, Dong, and Poibeau analyze arXiv papers at scale and identify previously unreported LLM-driven word usage shifts: increased frequency of 'beyond' and 'via' in titles and decreased frequency of 'the' and 'of' in abstracts. Current classifiers struggle to identify which specific LLM generated text in multi-class classification due to cross-model similarities. The findings provide empirical evidence of LLMs reshaping scientific writing at the corpus level.