PACE: New Method for Automated Algorithm Design via Primitive-Aware Code Evolution

A recent publication on arXiv (2608.07395) presents Primitive-Aware Code Evolution (PACE), a novel approach for automating algorithm design utilizing Large Language Models (LLMs). This study tackles a significant drawback in existing LLM methodologies, which often treat algorithms as single, unbreakable entities. This holistic view ties beneficial local logic to the entire program, resulting in the loss of useful code fragments when the program is discarded, complicating the evaluation of individual algorithm components. PACE separates local logic from complete programs by defining it as persistent units known as Executable Algorithmic Primitives (EAPs). It sustains a dynamic collection of EAPs to facilitate code-level transfer, with algorithm evolution guided by primitive-aware operators that ensure the preservation and transferability of these components. To effectively assess EAPs, PACE employs Thompson sampling based on parent-relative performance. This research, categorized as a cross announcement, is pertinent to artificial intelligence, machine learning, and automated algorithm design, potentially leading to the creation of more efficient and modular algorithms.