Beyond original detection scores: Markov information calibration improves machine-generated text detection

To address the problem that token-level detection scores in machine-generated text detection are susceptible to generation randomness, researchers propose a Markov information score calibration method, which uses a Markov random field to model the neighbor similarity and initial instability of contextual detection scores, and implements it as a lightweight component via mean-field approximation that can be seamlessly integrated into existing detectors. In experiments across various real-world scenarios such as cross-LLM and paraphrase attacks, the method significantly outperforms baselines with negligible computational overhead.