Teacher-Student Framework for Short-Context Classification with Domain Adaptation and Data Augmentation

Detecting AI-generated text is an important challenge, especially as large language models become better at creating content that looks like human writing. This paper presents a teacher-student framework to improve the performance and efficiency of short-context document classification. The framework uses domain adaptation and data augmentation to improve detection. The teacher model combines DeBERTa-v3-large and Mamba-790m models, using their fine-tuning on domain-specific data and semantic knowledge for accurate predictions. The student model works with short-context text of 128 and 256 tokens and learns from the teacher to balance accuracy and computational efficiency. To make the model more robust, we built a data generation and augmentation pipeline, applying techniques like spelling correction, character removal, and case flipping to increase data variety. The proposed framework outperforms current methods in accuracy and robustness, offering an efficient solution for detecting AI-generated text. This work also lays the groundwork for future studies on multilingual classification and real-time inference improvements in AI text detection.