PAPILLON introduces a novel approach for protecting user privacy when interacting with language models by intelligently delegating between local and cloud-based AI systems.

• Creates a privacy-conscious delegation framework that routes sensitive queries to local models and non-sensitive queries to more powerful cloud models
• Develops techniques to identify and protect personally identifiable information (PII) in user prompts
• Achieves superior privacy protection while maintaining high-quality responses through selective model routing
• Demonstrates practical implementation of privacy-preserving AI assistants without sacrificing capabilities

This research addresses critical security concerns for organizations using LLMs by preventing exposure of sensitive information to third-party providers while maintaining access to the most capable AI technologies.

PAPILLON: Privacy Preservation from Internet-based and Local Language Model Ensembles