P³Defer enables devices to intelligently decide when to use local AI versus sending data to more powerful cloud models while preserving user privacy.

• Creates a privacy-aware decision policy that determines when to use local vs. cloud AI
• Uses Chain-of-Thought enhancement to improve decision making accuracy
• Achieves better privacy-performance balance than existing cascade approaches
• Demonstrated across multiple domains including summarization and reasoning tasks

This research addresses critical security concerns in the growing field of on-device AI by ensuring sensitive data remains protected while still enabling access to powerful cloud capabilities when needed.

Privacy-preserved LLM Cascade via CoT-enhanced Policy Learning