Interpretable Multilingual Attribution of Frontier LLMs

Can you tell which frontier LLM wrote a piece of text, in any language, using features a person can actually inspect? We collected text from five frontier models across seven languages using content-matched prompts, then attributed it with interpretable stylistic features.

Character n-grams are more accurate within a single language, but the interpretable features explain their decisions and carry across languages. LLM judges asked to name the author stayed near chance. The paper was named a best paper finalist, one of 12 out of 130 accepted papers.

Paper available after the symposium (Dec 2026)