New Ctrl-G framework enhances LLM privacy with knowledge graphs

Researchers have developed Ctrl-G, a novel framework designed to enhance the privacy of knowledge graphs (KGs) when used with large language models (LLMs). This KG-RAG system addresses both structural and interaction exposure by creating query-conditioned, sanitized views of the graph locally before sending them to a remote LLM. Ctrl-G employs a hierarchical controller for planning and exploration, utilizing experience memory to reduce repeated disclosure across multiple queries. The framework demonstrates strong performance, achieving the highest accuracy on five out of six benchmarks and outperforming the strongest baseline by up to 13.8 points. AI

IMPACT Enhances privacy for LLM reasoning over knowledge graphs, potentially enabling more secure use of sensitive data.

RANK_REASON The cluster contains a research paper detailing a new framework for knowledge graph augmented LLM reasoning. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• Ctrl-G
• entity anonymization
• KG-RAG
• knowledge graph
• LLM
• relation-centric abstraction
• Web question answering with neurosymbolic program synthesis
• Xingyu Tang

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →