What Happens After the Shift? Training-Free Detection of LLM-Generated Text

The increasing use of large language models (LLMs) has created a growing need for reliable methods to distinguish human-written from LLM-generated text. Training-free detectors are attractive because they do not require labeled training data or access to the source generator, but existing approaches primarily rely on global likelihood statistics, temporal fluctuations, or spectral characteristics of token probability sequences. In this work, we investigate whether local transition behavior in token log-probability sequences provides an additional discriminative signal. We identify the dominant token-to-token transition and characterize the variability of the likelihood trajectory following this transition. Based on these observations, we propose ExpStabilityDet, a training-free detector that combines global likelihood statistics with transition-aware features, including the dominant transition magnitude and post-transition variability. We further introduce ExpStabilityDet++, a sampling-based extension that evaluates the observed text relative to conditionally generated contrast texts. Across multiple source models, datasets, response lengths, decoding strategies, and languages, our approach consistently demonstrates the usefulness of transition-aware likelihood characteristics. In the primary black-box evaluation, ExpStabilityDet achieves an average AUROC of 0.8877, while ExpStabilityDet++ achieves 0.9157, compared with 0.8785 and 0.8881 for SpecDetect and SpecDetect++, respectively. Additional experiments show that the proposed representation remains informative under cross-lingual and paraphrasing settings and provides a favorable performance–runtime trade-off. These results suggest that the relationship between a salient likelihood transition and its subsequent trajectory provides complementary information for training-free LLM-generated text detection.