Measuring Embedding Sensitivity to Authorial Style in French: Comparing Literary Texts with Language Model Rewritings
For researchers in authorship attribution and detection, this work provides insights into the persistence of stylistic signals in embeddings after LLM rewriting.
The study investigates how much stylistic information is encoded in language model embeddings and whether it persists after LLM rewriting, using French literary texts. Results show embeddings reliably capture authorial style and retain signals after rewriting, with LLM-specific patterns.
Large language models (LLMs) can convincingly imitate human writing styles, yet it remains unclear how much stylistic information is encoded in embeddings from any language model and retained after LLM rewriting. We investigate these questions in French, using a controlled literary dataset to quantify the effect of stylistic variation via changes in embedding dispersion. We observe that embeddings reliably capture authorial stylistic features and that these signals persist after rewriting, while also exhibiting LLM-specific patterns. These analytical results offer promising directions for authorship imitation detection in the era of language models.