<ul><li>Large Language Models (LLMs) rely on input prompts for performance optimization.</li><li>Existing prompt optimization work has focused on task-specific user prompts, neglecting system prompts applicable across tasks.</li><li>This study introduces bilevel system prompt optimization to create robust and transferable system prompts.</li><li>A meta-learning framework is proposed to optimize system prompts over various user prompts, enhancing adaptability to unseen tasks.</li></ul>

System Prompt Optimization with Meta-Learning

Discover more