EAVAE: Disentangling style from content for authorship attribution and AI text detection

EAVAE achieves SOTA on authorship attribution (Amazon Reviews, PAN21, HRS) and few-shot AI text detection (M4) by explicitly disentangling writing style from content. The framework uses separate encoders for style and content, pretrained with supervised contrastive learning and finetuned with a VAE, plus a discriminator that generates natural language explanations for its decisions. Code and datasets are open-source.

• #authorship attribution
• #ai-generated text detection
• #disentanglement
• #variational autoencoder
• #explainability