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

2026-06-17 · Source: Takara TLDR - Daily AI Papers · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, medium

Summary

SenFlow introduces a novel approach for sentence-level AI-generated text detection (S-AGTD) in hybrid documents, addressing limitations of existing methods that ignore inter-sentence dependencies and outdated benchmarks. The researchers developed MOSAIC, a new benchmark comprising 16,000 hybrid documents sourced from PubMed and XSum, generated by advanced LLMs like DeepSeek-V3.2 and Kimi K2, and filtered for perplexity consistency. SenFlow recasts S-AGTD as structured prediction, employing graph-based inter-sentence propagation combined with linear-chain CRF decoding in a single document-level pass. This method achieved state-of-the-art performance on MOSAIC, demonstrating a +4.15 pp average Macro-F1 margin in cross-domain transfer, the most challenging evaluation protocol. A significant finding is that even after perplexity filtering, AI-generated insertions exhibit a generator-dependent sentence-length disparity, which sentence-level detectors can still leverage for improved accuracy.

Key takeaway

For NLP Engineers developing AI-generated text detection systems, you should move beyond isolated sentence analysis. Incorporate inter-sentence flow modeling, like SenFlow's graph-based propagation and CRF decoding, to significantly improve accuracy in hybrid documents. Your benchmarks must include outputs from current LLMs such as DeepSeek-V3.2 and Kimi K2. Also, apply perplexity-consistency filters to ensure robustness against the latest generators. Consider analyzing subtle cues like sentence length variations.

Key insights

Detecting AI-generated text in hybrid documents requires modeling inter-sentence dependencies, not just isolated sentences, for improved accuracy.

Principles

• Inter-sentence dependencies improve S-AGTD.
• Benchmarks need current LLMs and quality filters.
• AI insertions show generator-specific length patterns.

Method

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

In practice

• Incorporate inter-sentence context for detection.
• Build benchmarks with recent LLMs and quality filters.
• Exploit generator-dependent sentence length gaps.

Topics

• AI-Generated Text Detection
• Hybrid Documents
• Inter-Sentence Flow Modeling
• Structured Prediction
• LLM Benchmarking
• DeepSeek-V3.2
• Kimi K2

Code references

• luojingkun22/SenFlow
• vinusankars/Reliability-of-AI-text-detectors

Best for: Research Scientist, AI Scientist, NLP Engineer, Machine Learning 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.
• Detecting Overflow in Compressed Token Representations for Retrieval-Augmented Generation — Token overflow in compressed LLM representations can be detected by incorporating query information into lightweight classifiers.
• Stanford AI Lab Papers and Talks at ACL 2022 — ACL 2022 showcased diverse NLP research from Stanford, spanning model pretraining, analysis, and evaluation.
• Data Machina #262 — New models and frameworks are advancing LLM efficiency, reasoning, and long-form content generation.
• From Raw Text to Intelligence: Building an NLP Pipeline Step by Step — NLP pipelines transform raw, unstructured text into numerical representations for machine learning models through systematic cleaning and vectorization.
• Do we really need to detect LLM-generated text? — Masked word prediction by LLMs can distinguish AI-generated text from human text without explicit training.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Takara TLDR - Daily AI Papers.