Segmenting Human-LLM Co-authored Text via Change Point Detection
Segmenting Human-LLM Co-authored Text via Change Point Detection
Authors: Mengchu Li, Jin Zhu, Jinglai Li, Chengchun Shi Date: 2026-05-05 Paper ID: openalex:2605.03723
Summary
This paper addresses the challenge of localizing human-written and LLM-generated segments within co-authored documents. By modeling the text authorship as a sequence with varying detection scores, the authors reframe the task as a change point detection problem. They propose two robust algorithms—a weighted version and a generalized version—to handle heterogeneous score distributions and demonstrate their theoretical minimax optimality and empirical superiority over baseline methods.
Key Contributions
- Formulates the task of segmenting human-LLM co-authored text as a change point detection problem in time-series analysis.
- Develops weighted and generalized change point detection algorithms specifically designed to account for heterogeneous score variability in LLM text detection.
- Provides theoretical proof establishing the minimax optimality of the proposed segmentation procedure.
Open Questions & Future Work
Archivist Review
The paper provides a statistical framework for text segmentation by reframing it as a change point detection problem. I have approved the open question regarding the bottleneck caused by the independence assumption, as it highlights a significant limitation in applying classical change point models to linguistically dependent co-authored sequences. The proposed algorithms were rejected as they represent specific methodological adaptations rather than broadly applicable novel mechanisms.
Approved Open Questions
- Co-authored Text Segmentation Dependency Bottleneck: The assumption of independence between sentence-level scores is a simplification that limits the applicability of current change point models in complex, cohesive prose, which is the primary domain of LLM-human co-authoring. Addressing these dependencies is crucial for improving boundary detection accuracy in real-world writing tasks.
Rejected Candidates
- [concept] Formulating authorship segmentation as change point detection (
formulation-of-authorship-segmentation-as-change-point-detection) - not_novel: Reframing a task using an existing statistical method is an application rather than a new reusable mechanism. - [concept] Weighted and generalized change point detection algorithms (
weighted-and-generalized-change-point-detection-algorithms) - subcomponent_of_broader_mechanism: These are specific algorithmic variations applied to a specific task, better categorized as methodological adjustments than a permanent concept.
Links
Metadata & Links
- url
- https://arxiv.org/abs/2605.03723
- paper_id
- 2605.03723
- paper_source
- openalex
- domain
- nlp
- tags
-
- architectures
-
- datasets
-
- concept_slugs
-
- dataset_slugs
-
- skill
- TimeSeriesSkill
- created_at
- 2026-05-08T05:44:30Z