Researchers have introduced a new benchmark called P3Bench to address personalized privacy control in large language models (LLMs). This benchmark extends contextual privacy policies to include user-specific disclosure preferences, recognizing that acceptable information disclosure can vary significantly between individuals even within the same context. Experiments revealed that prompt-based policies are unreliable for enforcing these personalized privacy settings, with models like Qwen2.5-7B and Gemma3-4B demonstrating high rates of policy ignorance. To overcome this, a novel inference-time method called Repair was developed, which uses attention head intervention to guide the LLM's disclosure behavior towards user-defined privacy policies. AI
IMPACT This research could lead to more robust LLM privacy controls, enabling safer deployment of agentic AI systems that handle sensitive user data.
RANK_REASON The cluster contains an academic paper detailing a new benchmark and method for LLM privacy. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
