LLM privacy reasoning improved via normative simulacra from fiction

A new arXiv preprint (2604.20904) proposes extracting normative simulacra—structured representations of norms and information flows—from fiction novels to fine-tune large language models (LLMs) for better privacy reasoning. The approach uses supervised learning followed by GRPO reinforcement learning, with a composite reward function combining programmatic signals (task clarity, structural completeness, internal consistency, context identification) and an LLM judge. This addresses the misalignment between LLM agents' information handling and users' contextual privacy expectations, as defined by the Contextual Integrity (CI) framework, without doubling inference costs or relying on narrow task-specific data.