Safety
SCOPE: Sequential Conformal Probing for Reliable OOD Rejection in LLM Services
The paper introduces SCOPE (Sequential Conformal OOD Probing and Evaluation), a novel framework designed to enhance out-of-distribution (OOD) rejection in large language model (LLM) services. SCOPE leverages a readable hidden layer and constructs a conformal gate with in-distribution calibration, utilizing a supermartingale e-process to certify service-boundary evidence. Experimental results demonstrate that SCOPE outperforms traditional final-layer detectors in gate-level rejection and provides insights into the geometric nature of OOD boundaries within hidden spaces, which is crucial for practitioners seeking reliable OOD handling in LLM applications.
oodllmrejection