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

Researchers introduce UH-NAS, an LLM-guided neural architecture search framework that optimizes neural networks for unconventional hardware platforms by co-designing for accuracy and hardware-specific constraints like energy efficiency and physical imperfections. The approach demonstrates superior performance on optical computing hardware compared to existing methods, advancing the practical deployment of AI on emerging computing substrates.

UH-NAS addresses a critical bottleneck in emerging computing: most neural architecture search methods are tightly coupled to specific hardware families, making them inflexible for the diverse landscape of unconventional processors entering the market. This research tackles the fundamental engineering challenge of co-optimizing AI models and hardware constraints simultaneously—a problem that becomes increasingly important as organizations explore optical, photonic, and specialized computing platforms beyond traditional GPUs and TPUs.

The motivation stems from the growing mismatch between standard deep learning optimization and real-world deployment requirements. Emerging hardware platforms introduce novel constraints including energy budgets, physical manufacturing tolerances, and precision limitations that conventional NAS algorithms ignore. By integrating language models as evolutionary operators and treating hardware as a pluggable backend, UH-NAS decouples the search algorithm from specific platforms while enabling fair cross-platform comparisons.

The framework's significance extends beyond academic interest. Organizations investing in optical computing, neuromorphic chips, and quantum-classical hybrid systems need design methodologies that account for hardware realities from the start. The demonstrated robustness under non-idealities suggests practical applicability rather than theoretical novelty. This approach reduces engineering cycles and accelerates the viability of alternative computing paradigms by automating architecture discovery.

Looking forward, the integration of LLMs into hardware-aware design workflows represents a broader trend in AI infrastructure. As hardware diversity expands, automated co-design frameworks become essential infrastructure. The next critical milestone involves whether such methods can scale to larger models and more complex hardware constraints, and whether practitioners will adopt LLM-guided NAS for production systems.

• →UH-NAS enables hardware-agnostic neural architecture search by treating hardware platforms as swappable backends with configurable constraints.
• →Language models function as evolutionary operators within the search process, improving diversity and robustness of discovered architectures.
• →The framework demonstrates superior performance on optical computing hardware compared to baseline NAS and existing LLM-to-NAS approaches.
• →Physical non-idealities and manufacturing tolerances are integrated into the search process, improving real-world deployment reliability.
• →The approach enables fair system-level comparison across heterogeneous hardware platforms without algorithm modification.