LLM-Guided Neural Architecture Search for Robust Co-Design of Physical Neural Networks

A new framework called UH-NAS aims to make it easier to design neural networks for unconventional computing hardware, treating the hardware itself as a swappable component in the search process, according to research submitted on 9 June 2026 [1][2]. The framework, detailed in a paper submitted to arXiv, is designed to address a core challenge in deploying neural networks on emerging platforms: the need to co-optimize a model's task accuracy with hardware-specific constraints like energy consumption and physical imperfections [2]. Neural networks, which are computational models composed of layers of connected artificial neurons, typically require significant computational resources for training and deployment [3]. The field of artificial intelligence has seen cycles of progress, with a major acceleration after 2012 when graphics processing units were widely adopted to speed up neural network training [4]. Machine learning, the broader discipline, focuses on algorithms that learn from data without explicit programming [5]. Existing neural architecture search (NAS) methods are often built for a single hardware family, which limits the ability to compare performance across different platforms [2]. The UH-NAS framework takes a different approach. It is hardware-agnostic and uses large language models (LLMs) as evolutionary operators to guide the search [1]. LLMs are a type of neural network trained on vast text corpora for tasks like language generation and are the foundation of modern chatbots [11]. By integrating them into the search process, UH-NAS co-optimizes for both accuracy and inference energy [2]. The system exposes the target hardware as a swappable backend, complete with per-platform energy models, physical constraints, and simulators for non-idealities [1]. This design allows for fair, system-level comparisons across different backends without altering the core search algorithm [2]. When tested on optical MZI hardware, UH-NAS discovered architectures that were more diverse and robust than those found by conventional baselines, and it outperformed existing approaches that use LLMs for neural architecture search [1]. The researchers also conducted additional studies on how the discovered architectures hold up under hardware non-idealities and examined the role of system prompts, underscoring the importance of co-design for new computing platforms [2]. The paper was posted on arXiv, an open-access repository for electronic preprints that has hosted over two million articles since its launch in 1991 and is not peer-reviewed [9]. The work appears within arXiv's machine learning section and is accessible through the arXivLabs framework, a community innovation space that hosts experimental tools and features on the site [1][8].