AUTO: LLM-Driven Strategic Search Optimizes GPU Code

AUTO, a novel framework, utilizes large language models (LLMs) to streamline design optimization, particularly in scenarios where defining search spaces and design parameters poses challenges. This framework distinguishes between high-level planning, conducted by a Strategist agent, and low-level execution, managed by concurrent Implementor agents, which iteratively enhance designs using explore-exploit techniques. When applied to three GPU code optimization tasks, AUTO surpasses in-lab-optimized code for chemical kinetics by as much as 1.74× for problem sizes reaching 10^5 cells and achieves up to 94% of cuBLAS double-precision performance in matrix multiplication. The research can be found on arXiv under the identifier 2511.22651, categorized as replace-cross.