Latent Trajectory Discrimination for AI-Generated Text Detection

arXiv cs.CL · 6 hours ago

Researchers propose GTCL, a method for detecting AI-generated text by analyzing how textual representations evolve sequentially through a document's latent space rather than treating text as static. The framework segments documents into ordered units, encodes them in embedding space, and applies contrastive learning to learn patterns of autoregressive generation. Testing on three benchmarks shows GTCL outperforms existing detection approaches by explicitly modeling sequential dynamics across different AI models and domains.