Multi-level contextual token relationship modeling for machine-generated text detection

The researchers propose a multi-level contextual token relationship modeling framework for machine-generated text detection, addressing the difficulty that token-level detection scores are susceptible to interference from generation randomness. The framework uses a lightweight Markov calibration module to correct token-level evidence and model local relationships before aggregation, uses explicit logical rules based on contextual score statistics for global rule-supported reasoning, and jointly fuses the two through multi-level inference. Experiments show that it achieves significant improvements in cross-LLM and cross-domain scenarios with low computational overhead.