Stemma: A Black-Box Method for LLM Provenance Testing

A recent publication on arXiv presents Stemma, a novel black-box technique for assessing the provenance of large language models (LLMs). This method overcomes the shortcomings of current techniques that depend on response-level traits, which may change during adaptation or deployment. By mapping open-ended outputs into a finite decision space, Stemma creates induced decision regions, minimizing surface-form variations. This approach redefines provenance testing to focus on the inheritance of decision regions. Empirical findings indicate that related models maintain source-induced regions more effectively than unrelated ones. Stemma also establishes stability, robustness, and specificity as complementary principles for selecting probes, ensuring reliable fingerprinting.