HLPD: Aligning LLMs to Human Language Preference for Machine-Revised Text Detection

The introduction of Human Language Preference Detection (HLPD) marks a significant advancement in the field of AI text detection. As large language models (LLMs) become more sophisticated, the challenge of identifying machine-generated content grows, particularly when revisions are made in a multi-task setting. HLPD addresses this by aligning detection models with human language preferences through a reward-based optimization process. This approach has demonstrated a 15.11% improvement in AUROC over the previous ImBD model and surpasses Fast-DetectGPT by 45.56%. Such enhancements are crucial in the ongoing battle against misinformation, as they enable more reliable identification of AI-generated texts, thus preserving the integrity of information in an increasingly digital world.