Latent Trajectory Discrimination for AI-Generated Text Detection

Source

• Source name: arXiv
• Source kind: paper
• Authors / organization: Gianluca Bonifazi, Christopher Buratti, Michele Marchetti, Federica Parlapiano, Giulia Quaglieri, Davide Traini, Domenico Ursino, Luca Virgili
• Venue: arXiv: cs.CL
• Published: 2026-07-16
• Collected: 2026-07-17
• URL: http://arxiv.org/abs/2607.14967v1
• PDF: https://arxiv.org/pdf/2607.14967v1
• DOI: UNKNOWN

Plain-Language Brief

In this paper, the authors propose a new approach for detecting AI-generated text called Geometric Trajectory and Contrastive Learning (GTCL). Unlike existing methods that focus on aggregate statistics or global compressions of texts, GTCL models how textual representations evolve across sequences. The method outperforms detection baselines consistently across three different benchmarks and various approaches.

What the Document Says

• Most AI-generated text detection (AIGTD) methods consider documents as static objects, using aggregate statistics or global compressions of the texts for decision-making.
• The authors argue that this approach overlooks the dynamic nature of autoregressive generation, where content evolves progressively through the latent space.
• In their paper, they reformulate AIGTD as a problem of distinguishing between latent generation trajectories.
• They propose Geometric Trajectory and Contrastive Learning (GTCL), a framework that segments documents into ordered local units, encodes each unit in an embedding space, and constructs a structured and sequence-level representation.
• GTCL applies contrastive learning to these trajectories to learn geometric regularities associated with autoregressive generation.
• Evaluations on three benchmarks show that GTCL outperforms detection baselines consistently across various approaches.
• The authors suggest that modeling trajectory differences could improve detection and open up a dynamic direction in AIGTD literature.

Key Concepts Mentioned

• AI-generated text
• AIGTD
• autoregressive generation
• latent space
• Geometric Trajectory and Contrastive Learning (GTCL)
• contrastive learning
• benchmarks

Claims Made by the Document

• Existing AIGTD methods overlook the dynamic nature of autoregressive generation.
• GTCL outperforms detection baselines consistently across various approaches and benchmarks.

Evidence Used

• The authors state that most existing AIGTD methods consider documents as static objects, while they argue that this approach overlooks the dynamic nature of autoregressive generation.
• Evaluations on three benchmarks show that GTCL outperforms detection baselines consistently.

Methods / Approach

UNKNOWN

Limitations Stated by the Authors

UNKNOWN

Limitations Noted by the Agent

Agent interpretation — not from the source.

UNKNOWN

Open Questions

UNKNOWN

Notes for Future Q&A

May include interpretation. Labeled where it does.

UNKNOWN

Raw Abstract / Excerpt

Most existing approaches to AI-Generated Text Detection (AIGTD) treat documents as static objects and base their decisions on aggregate statistics or globally compressed embeddings. However, this perspective overlooks the inherently dynamic nature of autoregressive generation, where content evolves progressively through the latent space. In this paper, we reformulate AIGTD as the problem of distinguishing between latent generation trajectories. Instead of relying on static representations, we model how textual representations evolve across the sequence. To this end, we propose Geometric Trajectory and Contrastive Learning (GTCL), a framework that segments the document into ordered local units, encodes each unit in an embedding space, and constructs a structured and sequence-level representation. GTCL then applies contrastive learning to these trajectories to learn geometric regularities associated with the autoregressive generation. Evaluations performed on three different benchmarks and several approaches show that GTCL outperforms detection baselines consistently, which implies that explicitly modeling sequential dynamics provides robust discriminative signals across models and domains. These results suggest that modeling trajectory differences could improve detection and open up a dynamic direction that has been underexplored in previous AIGTD literature.