Stemma: Induced Decision Regions Reveal LLM Provenance
Stemma tests model lineage by checking whether related LLMs preserve the same induced decision regions.
The paper maps open-ended model outputs into a finite decision space to reduce reliance on surface-level response patterns. Its experiments report that related models preserve a source model’s induced regions more strongly than unrelated ones. Across 770 source-suspect pairs from 56 public checkpoints, Stemma reaches 0.967 AUC and 87.8% TPR at 1% FPR. On 1,260 pairs across 91 deployment instances, it reports 0.995 AUC and 93.5% TPR at 1% FPR. ArXiv · AI/CL/LG's note
The paper maps open-ended model outputs into a finite decision space to reduce reliance on surface-level response patterns. Its experiments report that related models preserve a source model’s induced regions more strongly than unrelated ones. Across 770 source-suspect pairs from 56 public checkpoints, Stemma reaches 0.967 AUC and 87.8% TPR at 1% FPR. On 1,260 pairs across 91 deployment instances, it reports 0.995 AUC and 93.5% TPR at 1% FPR. ArXiv · AI/CL/LG's note
score 5