As the prompting capabilities of LLMs evolve, Tobias Schnabel and Jennifer Neville present SAMMO—a structure-aware approach that treats prompts as programs, applying compile-time optimizations to meta prompt programs. SAMMO boosts performance and generalizes beyond traditional methods, proving its validity across diverse tasks like instruction tuning, RAG pipeline tuning, and prompt compression across various LLMs.

Highlights:

• A structure-aware, compile-time optimization framework
• Encompasses a rich set of transformations for optimization
• Applicability and improvements demonstrated across multiple LLMs
• Open-source code available to facilitate further progress

Find the groundbreaking paper on prompt optimization on Arxiv.