Feature-Augmented Transformers for Robust AI-Text Detection Across Domains and Generators

Mohamed Mady, Johannes Reschke, Björn Schuller · 2026

Read Feature-Augmented Transformers for Robust AI-Text Detection Across Domains and Generators by Mohamed Mady, Johannes Reschke, Björn Schuller (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-03969.txt.

Description

Mohamed Mady, Johannes Reschke, and Björn Schuller study how well supervised detectors can tell AI-generated text from human writing when test data shift across domains and generators. They train transformer-based detectors on the HC3 PLUS dataset. They set a single decision threshold, chosen for best balanced accuracy on held-out validation. That threshold is then kept fixed for all downstream test distributions. The fixed threshold reveals domain- and generator-dependent error gaps under shift.

Base models reached near-ceiling in-domain performance. But performance under shift was brittle and strongly model-dependent. Feature augmentation via attention-based language feature fusion improved transfer. The best model reached about eighty-six percent balanced accuracy on the M4 benchmark. It beat strong zero-shot baselines by up to seven points. Multi-seed experiments confirmed high stability. Category-level breakdowns showed that readability and vocabulary features contribute most to robust transfer.

Catalog record

CREATORMohamed Mady, Johannes Reschke, Björn Schuller
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE46 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.03969

Browse the full library · About the Tower of Babel