SenFlow: Inter-Sentence Flow Modeling for AI-Generated Text Detection in Hybrid Documents

2026-06-17 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Natural Language Processing · Depth: Expert, quick

Summary

SenFlow is a novel method designed for sentence-level AI-generated text detection (S-AGTD) in hybrid documents, where human and LLM content co-exist. It addresses limitations of prior approaches that classify sentences in isolation and use outdated benchmarks. To facilitate this, the authors constructed MOSAIC, a new benchmark comprising 16,000 hybrid documents derived from PubMed and XSum, generated by DeepSeek-V3.2 and Kimi K2. MOSAIC incorporates a stringent perplexity-consistency filter, a feature absent in previous benchmarks. SenFlow recasts S-AGTD as a structured prediction problem over document sentence sequences, integrating graph-based inter-sentence propagation with linear-chain CRF decoding in a single document-level pass. This approach achieves state-of-the-art performance on MOSAIC, demonstrating a +4.15 pp average Macro-F1 margin on cross-domain transfer, the most challenging protocol. The research also found that even with perplexity filtering, AI insertions maintain a generator-dependent sentence-length gap.

Key takeaway

For NLP engineers developing AI-generated text detection systems, recognize that isolated sentence analysis is insufficient for hybrid documents. Your models should incorporate inter-sentence dependencies, as demonstrated by SenFlow's structured prediction approach, to achieve higher accuracy. Consider evaluating your detectors against the MOSAIC benchmark, which includes recent LLM outputs and a perplexity-consistency filter, to ensure robustness against modern generators.

Key insights

SenFlow improves AI-generated text detection in hybrid documents by modeling inter-sentence dependencies and using a new benchmark.

Principles

• Inter-sentence dependencies are crucial for S-AGTD.
• Structured prediction enhances sentence-level detection.
• Perplexity filters don't eliminate all AI cues.

Method

SenFlow recasts S-AGTD as structured prediction, integrating graph-based inter-sentence propagation with linear-chain CRF decoding in a single document-level pass over a sentence graph.

In practice

• Use MOSAIC benchmark for S-AGTD evaluation.
• Consider inter-sentence context for detection.
• Analyze sentence length as an AI insertion cue.

Topics

• AI-Generated Text Detection
• Hybrid Documents
• Structured Prediction
• Inter-Sentence Dependencies
• MOSAIC Benchmark
• Large Language Models

Code references

• luojingkun22/SenFlow

Best for: Research Scientist, AI Scientist, NLP Engineer

Related on AIssential

• SenFlow: Inter-Sentence Flow Modeling for AI-Generated Text Detection in Hybrid Documents — Detecting AI-generated text in hybrid documents requires modeling inter-sentence dependencies, not just isolated sentences, for improved accuracy.
• SV-Detect: AI-generated Text Detection with Steering Vectors — SV-Detect uses steering vectors from frozen LMs to robustly detect AI-generated text, even under distribution shifts.
• Operation-Guided Progressive Human-to-AI Text Transformation Benchmark for Multi-Granularity AI-Text Detection — AI-text detectability is non-monotonic, influenced by edit operations and revision history, not just AI content proportion.
• An End-to-End Hybrid Framework for Rumour Detection in Low-Resources Algerian Dialect — A hybrid framework combining transformer embeddings with classical classifiers enables effective rumour detection in low-resource Algerian dialect.
• A Structured Benchmark for Text-Guided Anomaly Detection: When Language Stops Conditioning the Decision — Current multimodal anomaly detection systems exhibit superficial text guidance, often relying more on visual features than language conditioning.
• Enhancing SDG-Text Classification with Combinatorial Fusion Analysis and Generative AI — Combinatorial Fusion Analysis (CFA) significantly enhances SDG text classification by combining diverse models and synthetic data.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.