Latent Trajectory Discrimination for AI-Generated Text Detection

GTCL detects AI-generated text by modeling the geometric evolution of embeddings across a sequence rather than using aggregate statistics. The framework segments documents into ordered local units and employs contrastive learning to identify regularities in latent trajectories characteristic of autoregressive generation processes.