The arXiv preprint "Contextual Candour: Role-Aware Graduated Disclosure for LLMs over Mixed-Sensitivity Contexts" tackles a growing problem: large language models increasingly operate over contexts that mix public and confidential information. The authors argue that models must reason across this range while disclosing only what the user is authorised to receive.
The proposed approach, Contextual Candour, is described as role-aware and graduated. Rather than treating access as a simple binary, it appears to tailor disclosure based on the user's role and the sensitivity of each piece of information. The abstract contrasts this with existing approaches that enforce access restrictions, though the full details of those methods are not included in the summary.
The significance lies in moving from blanket access control toward more nuanced, role-aware disclosure. This could be relevant for enterprise or government settings where users with different clearance levels share the same context. However, the abstract does not provide experimental results or implementation specifics, so the practical effectiveness of the framework remains to be demonstrated.