LLMENAS: Evolutionary Neural Architecture Search via Large Language Model Guidance

Differentiable Neural Architecture Search (NAS) and traditional evolutionary approaches frequently struggle with premature convergence to local optima. To overcome this limitation, we propose LLMENAS, a hierarchical framework that introduces trajectory-aware fitness design as the upper-level optimizer. By analyzing the convergence state of historical optimization trajectories, the Large Language Model (LLM) acts as a fitness designer to dynamically design fitness functions. This mechanism enables the search to navigate complex landscapes and escape local optima. Furthermore, we introduce a closed-loop self-improving mechanism, enabling the LLM to iteratively enhance its design strategies through self-reflection and self-refinement based on feedback. Extensive experiments show that LLMENAS achieves competitive results, with top-1 accuracies of 97.58% on CIFAR-10, 83.52% on CIFAR-100, and 75.6% on ImageNet-1k. Furthermore, it is achieved with remarkable efficiency, costing only 0.15 GPU days on the CIFAR 10 datasets and 2 GPU days on ImageNet. The source code is publicly available at: https://github.com/LLMENAS/LLMENAS.