Symbolic Prompt Program Search: A Structure-Aware Approach to Efficient Compile-Time Prompt Optimization

Source: DOI Authors: Tobias Schnabel, Jennifer Neville Date: 2024

Abstract / key passage

This paper introduces SAMMO, a framework for symbolic prompt-program search aimed at compile-time optimization of reusable prompt programs. Instead of assuming a flat prompt string or a fixed prompt structure, SAMMO represents prompt programs symbolically and searches over structure-aware transformations. The reported gains cover instruction tuning, RAG pipeline tuning, and prompt compression across several language models.

Harness takeaway

SAMMO is unusually relevant to harness design because it treats prompt programs as versioned artifacts that can be optimized before deployment. That is much closer to real control-plane engineering than endlessly editing one monolithic prompt string.