How Semantically Stable Are LLM Refusals? Measuring Confusion in Local Safety Boundaries
This work introduces Semantic Confusion, a failure mode that captures contradictory refusal decisions across meaning-preserving paraphrases and proposes three model-agnostic token-level metrics: Confusion Index, Confusion Rate, and Confusion Depth.
Riad Ahmed Anonto, Md Labid Al Nahiyan, Md Tanvir Hassan
· 2 citations