Tag
This paper proposes a multi-level contextual token relation modeling framework for machine-generated text detection, integrating local Markov-informed calibration and global rule-support reasoning to improve detection across cross-LLM and cross-domain settings with low computational overhead.