The 9 October arXiv preprint examines formatting shortcuts in selected benchmarks that distinguish human-written from machine-generated scientific text. The authors trained classifiers on surface features that do not use the words themselves across ten dataset subsets, and report that high scores on arXiv subsets can partly reflect collection differences such as line breaks retained during PDF extraction. The first page lists co-author Andrey Grabovoy’s affiliation as Moscow State University, Moscow, Russia.

In the authors’ experiments, four formatting features reached an AUC of 0.943–0.997 on three arXiv subsets; normalizing text formatting lowered the surface probe’s macro-F1 by 0.199–0.268 points on those subsets. They also propose counterfactual edits that change formatting without changing words. The findings are limited to the selected datasets and detectors studied; they do not show that every detector uses these cues or that distinguishing human from machine text is impossible. The study’s practical point is that benchmark results should be checked for formatting differences introduced during data preparation.