LLM Alignment--Utility Asymmetry under Semantic-Preserving Transformations
Synthetic semantic-preserving transformations that are rule-based and invertible are used as a probe of alignment generalization and suggest that semantic-preserving distribution shifts can expose a recurring gap in how utility and alignment generalize in current LLMs.
Mo-Han Li, Cheng-Yu Yu, Francesco Sovrano et al.
· 0 citations