Validity-Aware Jailbreak Evaluation for Large Language Models
This work proposes Sequential Epistemic and Action-Level Validation (SEAV), a verification-centric jailbreak evaluation framework that decomposes responses into ordered steps and evaluates both validity and correctness, and shows that enforcing correctness substantially reshapes measured robustness.
Qilong Wu, Sahil Wadhwa, Pranab Mohanty et al.
· 0 citations