Reinforcing privacy reasoning in LLMs via normative simulacra from fiction

ArXi:2604.20904v1 Announce Type: new Information handling practices of LLM agents are broadly misaligned with the contextual privacy expectations of their users. Contextual Integrity (CI) provides a principled framework, defining privacy as the appropriate flow of information within context-relative norms. However, existing approaches either double inference cost via supervisor-assistant architectures, or fine-tune on narrow task-specific data.