LLMs and Contextual Integrity

Research on Large Language Models has led to a focus on contextual integrity, particularly in relation to the use of persistent memory from past interactions. This memory can enhance personalization and task performance, but it also poses risks when sensitive information is revealed in inappropriate contexts. A benchmark called CIMemories has been developed to evaluate whether LLMs can appropriately control information flow from memory. The goal of this benchmark is to assess the ability of LLMs to maintain contextual integrity. This is an important consideration as LLMs become increasingly prevalent and handle sensitive information. The development of CIMemories is part of a broader effort to understand and address the risks associated with LLMs and contextual integrity.