Project Page

RATIO: Robust AI Text Involvement Estimator

A two-stage detector for fine-grained AI involvement ratio estimation in mixed-authorship text.

Final Year Project ยท Beijing University of Technology and University College Dublin

Le Liu Yunhan Gao Ziheng Wang Sicheng Yi Bohan Zhang

Beijing University of Technology logo University College Dublin logo

Motivation

RATIO motivation figure showing the need for fine-grained AI involvement ratio estimation.

Method

RATIO first trains a proportion-aware detector on PACT mixed-authorship documents, then improves robustness with RCDPO training on clean and rewritten pairs.

Data

PACT provides controlled human-AI mixed text with document-level ratio labels and sentence-level provenance labels.

Paper: Replace this anchor after release.

PACT Construction Demo: The demo at uspa.zhangbh.com illustrates the dataset construction workflow used by PACT.