ML Security PapersHLPD: Aligning LLMs to Human Language Preference for Machine-Revised Text Detection
Fangqi Dai 1,2, Xingjian Jiang 1,3, Zizhuang Deng 1,4
2 Xi’an Jiaotong-Liverpool University
3 Chalmers University of Technology
4 State Key Laboratory of Cryptography and Digital Economy Security
Published on arXiv
2511.06942
Output Integrity Attack
OWASP ML Top 10 — ML09
Key Finding
HLPD achieves 15.11% relative improvement in AUROC over ImBD and 45.56% over Fast-DetectGPT when detecting GPT-series machine-revised text in black-box settings.
HLPD / HLPO
Novel technique introduced
To prevent misinformation and social issues arising from trustworthy-looking content generated by LLMs, it is crucial to develop efficient and reliable methods for identifying the source of texts. Previous approaches have demonstrated exceptional performance in detecting texts fully generated by LLMs. However, these methods struggle when confronting more advanced LLM output or text with adversarial multi-task machine revision, especially in the black-box setting, where the generating model is unknown. To address this challenge, grounded in the hypothesis that human writing possesses distinctive stylistic patterns, we propose Human Language Preference Detection (HLPD). HLPD employs a reward-based alignment process, Human Language Preference Optimization (HLPO), to shift the scoring model's token distribution toward human-like writing, making the model more sensitive to human writing, therefore enhancing the identification of machine-revised text. We test HLPD in an adversarial multi-task evaluation framework that leverages a five-dimensional prompt generator and multiple advanced LLMs to create diverse revision scenarios. When detecting texts revised by GPT-series models, HLPD achieves a 15.11% relative improvement in AUROC over ImBD, surpassing Fast-DetectGPT by 45.56%. When evaluated on texts generated by advanced LLMs, HLPD achieves the highest average AUROC, exceeding ImBD by 5.53% and Fast-DetectGPT by 34.14%. Code will be made available at https://github.com/dfq2021/HLPD.
Key Contributions
- Human Language Preference Optimization (HLPO): a reward-based alignment process that shifts a scoring model's token distribution toward human-like writing to improve machine-revised text detection
- Adversarial multi-task evaluation framework using a five-dimensional prompt generator and multiple advanced LLMs to create diverse revision scenarios (polish, expand, rewrite)
- HLPD achieves 15.11% relative AUROC improvement over ImBD and 45.56% over Fast-DetectGPT on GPT-series-revised texts in black-box settings
🛡️ Threat Analysis
Primary contribution is a novel AI-generated/machine-revised text detection architecture (HLPD/HLPO). The paper proposes a fundamentally new detection methodology — reward-based human language preference optimization — to distinguish human-written from machine-revised text, directly addressing output integrity and content provenance.