●researcherThis approach allows for scaling DP-protected inference to massive contexts which was previously computationally and utility-prohibitive.
●policyThis provides a technical mechanism for maintaining user privacy in LLM-as-a-Service models during long-context interactions.