| Author | Vladimir Chaikin |
| Consultant | German Gritsai |
| Advisor | Andriy Grabovoy, PhD |
This paper invistigates the problem of robust AI-generated text detection under distribution shift. While simple perplexity is a simple solution for such detection, its oppurtunities lose significantly on out-of-domain data. We propose a novel method - calibrated perplexity - which normalizes raw perplexity scores using domain-specific language model baselines. This approach constructs a stable feature space less sensitive to text length, genre, and language variations. Experiments across multiple domains and languages are going to demonstrate improved in-domain and out-of-domain detection accuracy compared to standard perplexity-based methods.
If you find our work helpful, please cite us.
@article{citekey,
title={Title},
author={Name Surname, Name Surname (consultant), Name Surname (advisor)},
year={2025}
}Our project is MIT licensed. See LICENSE for details.