In the field of AI-generated text detection (AIGTD), most existing approaches treat documents as static objects and base their decisions on aggregate statistics or globally compressed embeddings. This perspective overlooks the dynamic nature of autoregressive generation, where content evolves progressively through the latent space. This paper reformulates AIGTD as the problem of distinguishing between latent generation trajectories.
We propose the Geometric Trajectory and Contrastive Learning (GTCL) framework to model how textual representations evolve across the sequence. GTCL segments the document into ordered local units, encodes each unit in an embedding space, and constructs a structured sequence-level representation.
It then applies contrastive learning to these trajectories to learn geometric regularities associated with autoregressive generation.
Evaluations on three different benchmarks and various methods show that GTCL consistently outperforms detection baselines, indicating that explicitly modeling sequential dynamics provides robust discriminative signals across models and domains.
These results suggest that modeling trajectory differences could enhance detection and open a dynamic direction that has been underexplored in previous AIGTD literature.
Blogger's Review: This study offers a novel perspective on AI-generated text detection by dynamically modeling the variations in generation trajectories, significantly improving detection performance. This approach not only enriches existing techniques but also points to future research directions that deserve further exploration.