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

2026-06-18 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, extended

Summary

SenFlow introduces a novel approach to Sentence-level AI-Generated Text Detection (S-AGTD) for hybrid documents, where human and LLM content are interleaved. It addresses limitations of prior methods that classify sentences in isolation and use outdated benchmarks. The system recasts S-AGTD as structured prediction, integrating graph-based inter-sentence propagation with linear-chain Conditional Random Field (CRF) decoding. Alongside SenFlow, the MOSAIC benchmark was constructed, comprising 16,000 hybrid documents from PubMed and XSum, generated by DeepSeek-V3.2 and Kimi K2, and filtered for perplexity consistency. SenFlow achieves state-of-the-art performance on MOSAIC, demonstrating a +4.15 pp average Macro-F1 margin on cross-domain transfer. A key finding is that even after perplexity equalization, AI insertions retain a generator-dependent sentence-length gap that S-AGTD can exploit.

Key takeaway

For Machine Learning Engineers developing AI text detectors, current sentence-level methods often fail on hybrid documents. You should integrate inter-sentence flow modeling and structured prediction, as demonstrated by SenFlow, to significantly improve detection accuracy and generalization across unseen domains and generators. This approach offers a robust solution for identifying interleaved human and AI content, even when perplexity cues are minimized, providing a more reliable tool for content authenticity.

Key insights

Modeling inter-sentence dependencies via structured prediction significantly enhances AI-generated text detection in hybrid documents.

Principles

• Authorship transitions often cluster in contiguous blocks.
• Semantic similarity reveals long-range authorship cues.
• Perplexity filtering does not eliminate all AI detection signals.

Method

SenFlow extracts features from a LoRA-aligned Llama-3.1-8B-Instruct proxy, fuses semantic and style embeddings using cross-attention, then refines representations with a Hybrid Adjacency GCN and decodes with a linear-chain CRF for joint sequence prediction.

In practice

• Integrate graph-based inter-sentence modeling for S-AGTD.
• Construct benchmarks with perplexity-consistency filters.
• Analyze sentence length as a detection feature.

Topics

• AI-Generated Text Detection
• Hybrid Documents
• Sentence-Level Detection
• Graph Neural Networks
• Conditional Random Fields
• LLM Benchmarking
• DeepSeek-V3.2

Code references

• luojingkun22/SenFlow

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• SenFlow: Inter-Sentence Flow Modeling for AI-Generated Text Detection in Hybrid Documents — SenFlow improves AI-generated text detection in hybrid documents by modeling inter-sentence dependencies and using a new benchmark.
• 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.
• Sem-Detect: Semantic Level Detection of AI Generated Peer-Reviews — AI-generated peer reviews converge semantically, while human reviews, even LLM-refined, retain unique ideas.
• 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.
• The PDF Paradox: Why Document Parsing Is Still Hard — And Why the Hybrid Stack Is Winning — PDF's fixed-layout design, while excellent for rendering, creates a significant parsing challenge for AI.
• Text-as-Signal: Quantitative Semantic Scoring with Embeddings, Logprobs, and Noise Reduction — Text can be transformed into operational, continuous semantic data points for AI engineering via embeddings and logprob scoring.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.