Stemma Fingerprints LLM Lineage From Induced Decision Regions: 0.967 AUC Across 770 Source-Suspect Pairs

arXiv 2607.25880·medium signalXBlueskyLinkedInCopy link

Stemma (arXiv 2607.25880, 2026-07-28) reframes LLM provenance testing away from response-level characteristics — which shift under adaptation and deployment even when meaning is unchanged — by mapping open-ended outputs into a finite decision space and measuring whether a suspect model inherits the source's induced decision regions. Using stability, robustness, and specificity as complementary probe-selection principles, it achieves 0.967 AUC and 87.8% TPR at 1% FPR across 770 source-suspect pairs drawn from 56 public checkpoints spanning diverse weight transformations, and 0.995 AUC with 93.5% TPR at 1% FPR on 1,260 pairs covering 91 deployment instances. That substantially outperforms four representative baselines and gives model owners a black-box tool for detecting unauthorized derivatives.