AgentNAS Combines LLMs and NAS for Superior AI Architectures

Traditional Neural Architecture Search (NAS) methods, while efficient, are constrained by manually defined search spaces that demand significant domain expertise and task-specific rebuilding. Researchers have introduced AgentNAS, a novel mechanism that integrates the open-ended design capabilities of Large Language Models (LLMs) with the systematic search power of NAS. AgentNAS begins with an LLM generating a high-quality initial architecture, which it then decomposes into a "slotted architecture" – a scaffold with interchangeable module slots. This scaffold automatically defines a bounded, task-specific search space for conventional NAS to explore without manual engineering. AgentNAS, structured as a three-phase pipeline, demonstrated superior performance across 17 diverse tasks, including classification, regression, and segmentation, spanning various modalities. It achieved new state-of-the-art results on 11 of these tasks, surpassing even expert-designed baselines. Ablation studies confirmed the complementary nature of the two search mechanisms: the LLM-generated seed alone often outperformed baselines, and NAS further enhanced gains through combinatorial recombination, a capability LLMs cannot replicate independently. This robust division of labor holds true across different LLM capabilities.