Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This paper introduces MultiGhostBench, a comprehensive multilingual benchmark designed to evaluate long-form LLM-generated text attribution across various languages and scripts, featuring 928 books generated by five recent LLMs. The benchmark reveals that existing authorship attribution methods struggle with consistent performance under distribution shifts, highlighting a significant gap in current approaches, particularly in multilingual contexts. Notably, while transformer-based detectors can maintain generator-related information across languages, their effectiveness varies significantly, emphasizing the need for more robust attribution methods in diverse linguistic settings.
No authorship attribution method reliably outperforms others across multilingual contexts, revealing critical vulnerabilities in current approaches.
While existing work on LLM authorship attribution (AA) has made progress, available benchmarks remain limited, often focusing on English, controlled settings, or relatively outdated models, with the few multilingual studies considering only relatively short texts. We introduce MultiGhostBench, a multilingual benchmark comprising 928 books generated by five recent LLMs across six languages and three scripts, with an average length of approximately 59K words per book. The benchmark supports evaluation under domain, author, and language shifts. Evaluation of representative AA methods shows that no single method consistently performs best across settings, and performance generally degrades under distribution shifts. Transformer-based detectors can retain generator-related information across languages, although transfer effectiveness varies by language pair, whereas statistical and fingerprint-based detectors are more language-dependent. We envision MultiGhostBench as a valuable resource for the development and evaluation of robust AA methods. The dataset and code can be found at https://github.com/GrecoMT/MultiGhostBench.