Tag
This paper introduces the concept of the audit gap between behavioral safety and representation-level robustness in LLMs, proposing an intervention-based evaluation framework and the Latent Vulnerability Score (LVS) to measure hidden vulnerabilities.