LLM-NAS: LLM-driven Hardware-Aware Neural Architecture Search

Large language models (LLMs) are a promising tool for neural architecture search (NAS), but direct prompting in hardware-aware NAS often leads to biased exploration, with candidate architectures concentrated in limited complexity regions. We propose LLM-NAS, a framework that improves Pareto-front coverage by partitioning the search space into complexity-based niches and updating the prompt context with search feedback across generations. Within each niche, the LLM acts as a search operator to generate increasingly informative candidates, while predictor-based accuracy estimation and latency lookup provide efficient training-free evaluation. Experiments on hardware-aware NAS benchmarks show that LLM-NAS achieves stronger Pareto quality than competitive baselines in terms of hypervolume (HV) and inverted generational distance (IGD), while maintaining low online search overhead. Ablation results further show that both complexity-driven partitioning and prompt co-evolution are important to the final performance.