Reproduce “SeqXGPT: Sentence-Level AI-Generated Text Detection” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Pengyu Wang, Linyang Li, Ke Ren, Botian Jiang, Dong Zhang, Xipeng Qiu
Researchers at Fudan University developed SeqXGPT, a method for sentence-level AI-generated text detection that extracts word-wise log probabilities from multiple white-box LLMs and processes them with a hybrid CNN-Transformer architecture. It achieved over 95% Macro-F1 in mixed-model detection tasks and demonstrated robust generalization on out-of-distribution datasets.
Key insights
- Word-wise log probabilities from multiple LLMs, when aligned and processed as sequential numerical features, contain highly discriminative information for identifying AI-generated content.
- A hybrid architecture combining Convolutional Neural Networks (CNNs) for local feature extraction and Transformer layers for capturing long-range dependencies effectively processes these unique "wave-like" log probability inputs.
- Fine-grained sentence-level AIGT detection requires a fundamentally different approach than document-level methods, particularly benefiting from sequence labeling and robust feature engineering on short text segments.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.