Official Review · ICLR 2024

The authors propose a diffusion based model for neural architecture generation. By using the guidance from the predictor, DiffusionNAG can generate task-specific architectures. They verify the effectiveness of DiffusionNAG on Transferable NAS and Bayesian Optimization-based NAS. This paper is well-motivated. Previous work only uses the diffusion model to model undirected graphs for neural architecture generation. Using the diffusion model for modeling the directed graph is needed. Moreover, classifier guidance is well suited for the Transferable NAS and substitute step 3) in BO-based NAS approaches, where the dataset and the pre-trained classifier are given as the setup. 1."Whether diffusion model for neural architecture generation is more efficient than Mutation + Random" is a good question. It is important to ask "when they are better". * The strength of the predictor guidance is an important hyperparameter. In this paper, a good performance can be achieved by cross-validation. Could the author provide some ablation studies of how this hyperparameter affects the performance? How could we effectively define the search range for cross-validation? * Will using the diffusion model for neural architecture generation have some model collapse phenomenon? Does the Diffusion model only generate some similar architecture, and achieve the better "worse case accuracy" by sacrificing the diversity? If that is the case, any explore and exploit framework for it? Please see the Weaknesses part.