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

A new detection method called SenFlow can identify AI-generated sentences inside documents co-written by humans and large language models, outperforming prior approaches on a freshly constructed benchmark of 16,000 hybrid texts, according to research published on arXiv. The work addresses sentence-level AI-generated text detection (S-AGTD) in hybrid documents, where human and machine prose are interleaved. The authors argue that existing methods treat each sentence in isolation, ignoring dependencies between sentences, and that current benchmarks do not reflect the newest generation of language models [1]. To close those gaps, the team built MOSAIC, a benchmark comprising 16,000 hybrid documents drawn from PubMed and XSum and produced by two recent large language models: DeepSeek-V3.2 and Kimi K2 [1]. DeepSeek is a Hangzhou-based AI company founded in 2023 that develops open-weight large language models; its V3 model was reportedly trained for roughly US$6 million, a fraction of the cost of comparable Western models [5]. The researchers recast sentence-level detection as a structured prediction problem over the entire document sequence. Their proposed system, SenFlow, combines graph-based inter-sentence propagation with linear-chain conditional-random-field decoding in a single pass over a sentence graph [1]. The architecture draws on principles similar to those used in convolutional neural networks, which learn features by sliding filters across input data and have been widely applied to natural language processing tasks [3]. By modeling relationships between adjacent sentences rather than classifying each one independently, SenFlow captures contextual signals that single-sentence classifiers miss. On the MOSAIC benchmark, SenFlow achieved state-of-the-art results across three evaluation protocols of increasing difficulty. On the hardest protocol — cross-domain transfer — it posted an average Macro-F1 margin of +4.15 percentage points over the next-best method [1]. The benchmark was constructed under stringent quality controls, including a perplexity-consistency filter absent from earlier datasets, which equalizes overt statistical cues that simpler detectors might exploit [1]. Even after that equalization, the study found that AI-generated insertions retain a generator-dependent sentence-length gap that sentence-level detectors can still leverage [1]. The paper does not include quotes from the authors, and the research has not yet been peer-reviewed. Code and data have been made publicly available on GitHub [1].