Tag
This paper applies the likelihood ratio framework for forensic authorship attribution to Japanese texts, fusing stylometric features with embedding-based systems to improve discrimination and calibration.
Introduces READER, a lightweight framework for dynamic black-box LLM provenance that uses a frozen proxy LLM to extract authorship evidence from responses and performs Bayesian evidence accumulation across multiple queries, achieving high accuracy on the Agent500 dataset.
Introduces PromptPrint, a systematic study showing that users' habitual vocabulary and syntax in LLM prompts form a learnable behavioral biometric, with lexical features outperforming semantic encoders and revealing a uniqueness–consistency paradox.
Sem-Detect introduces a method to distinguish AI-generated peer reviews from human-written ones by combining textual features with claim-level semantic analysis. It achieves a 25.5% improvement in true positive rate at 0.1% false positive rate over baselines, and shows that LLM-refined human reviews retain distinct semantic signals, with fewer than 3.5% misclassified as AI-generated.
This paper uses mechanistic interpretability to explain why authorship attribution models fine-tuned with the same encoder, data, and loss can differ four-fold in performance depending on the scoring mechanism. It finds that the scorer determines where the encoder consolidates authorship signal, with mean pooling forcing early consolidation and late interaction allowing late consolidation.
EditLens is a regression model that quantifies the extent of AI editing in text, achieving state-of-the-art performance on binary and ternary classification tasks distinguishing human, AI, and mixed writing. It addresses the gap in detecting AI-edited rather than fully AI-generated text, with implications for authorship attribution, education, and policy.
A foundational study on applying stylometric authorship attribution to threat intelligence, using Japanese Rakuten reviews to compare TF-IDF+LR, BERT embedding, BERT fine-tuning, and metric learning methods. BERT-FT performed best overall, but TF-IDF+LR proved more stable and efficient when scaling to hundreds of authors.