Reproduce “Robust AI-Generated Text Detection by Restricted Embeddings” 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

Kristian Kuznetsov, Eduard Tulchinskii, Laida Kushnareva, German Magai, Serguei Barannikov, Sergey Nikolenko, Irina Piontkovskaya

Researchers at AI Foundation and Algorithm Lab and Skoltech developed methods to improve the robustness of AI-generated text detectors by selectively removing specific linear subspaces from Transformer model embeddings, achieving up to 14% higher out-of-distribution accuracy for encoder-based models. The study found that erasing global syntactic and semantic features from embeddings can enhance generalization, while also highlighting architectural differences in how encoders and decoders learn text representations.

Key insights

  • Simple linear transformations applied to Transformer embeddings, such as coordinate-based removal and attention head pruning, can significantly improve the out-of-distribution robustness of AI-generated text detectors, particularly for encoder-based models.
  • Encoder-based models (like RoBERTa and BERT) and decoder-based models (like Phi-2 and MiniCPM) exhibit fundamental differences in their embedding space geometry, making subspace removal effective for encoders but less so for decoders.
  • Removing global syntactic information (e.g., top-level syntactic structure, tree depth) and word semantics from embeddings consistently improved cross-domain generalization in ATD.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.