Large Language Model Assisted Adversarial Robustness Neural Architecture Search

Large Language Models (LLMs) have shown signif-icant promise as evolutionary optimizers. This paper introduces a novel LLM-based Optimizer (LLMO) to address Neural Archi-tecture Search Considering Adversarial Robustness (ARNAS), a classic combinatorial optimization problem. Using the standard CRISPE framework, we design the prompt and employ Gemini to iteratively refine solutions based on its responses. In our numerical experiments, we investigate the performance of LLMO on NAS-Bench-201-based ARNAS tasks with CIFAR-IO and CIFAR-IOO datasets. The results, compared with six well-known metaheuristic algorithms (MHAs), highlight the superiority and competitiveness of using LLMs as combinatorial optimizers. The source code is available at https://github.com/RuiZhong961230/LLMO.