Reproduce “MASH: Evading Black-Box AI-Generated Text Detectors via Style Humanization” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yongtong Gu, Songze Li, Xia Hu
MASH rewrites AI-generated text in a human-like style to test evasion of black-box detectors. It trains a BART-based paraphraser on AI–human text pairs, aligns it using detector feedback, and can refine sentence-level edits while checking detector judgments. Across six English datasets and five detectors, the authors report 92% average attack success, but Reuters cross-domain transfer is a major exception.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.