Official Review · ICLR 2025

This study improved DiffusionNAG by introducing a multi-objective approach which modifies DiffusionNAG's reverse diffusion process as a reverse diffusion guidance process. Other than accuracy, #params, MACs and inference latency are also considered in the multi-objective metrics. The proposed method POMONAG has been tested on NASBench201 and MobileNetV3 with 15 image classification tasks, showing better performance than DiffusionNAG and a series of other methods. The motivation of this study, introducing multi-objective evaluation in NAS, is commendable as a task in reality is often not just about accuracy. Other metrics should be considered simultaneously as well. The writing is easy to follow. It is nice to see equations with highlights of different colours. **First of all**, the work claims to be on Pareto multiobjective search for architectures. However, that point is not obvious from the paper. * What are the benefits of using the proposed POMONAG? * How can a Pareto front be generated and utilized? Need to explicitly demonstrate how POMONAG generates and utilizes Pareto fronts. * How can users select architectures from the Pareto front according to their needs or under different circumstances? Show examples of such selection based on different priorities, for example prioritizing small-size architectures for portable devices or focusing on latency reduction etc. * It seems non-dominated sorting is absent. Explain how non-dominated sorting is incorporated or can be incorporated in POMONAG. * In its current form, the paper reads like a combination or integration of single-objective evaluations rather than a multi-objective evaluation. The equation of POMONAG at Line 209/210 is a linear combination of four objectives. Please clarify if the linear combination of objectives is intended as a scalarization approach. If so, discuss its limitations. --- **Secondly**, the performance of POMONAG appears better than DiffusionNAG and other methods shown in the paper. However many SOTA methods, especially zero proxy methods are missing. Their reported performance is similar or even better, for example, SWAP-NAS by Peng et al, ICLR'24, ZiCo by Li et al, ICLR'23, MeCo by Jiang et al, NeurIPS'23. * Include a comparison with these SOTA methods. If a direct comparison is not possible, explain why and discuss the limitations of the current evaluation. * Discuss how POMONAG's approach differs from or improves upon zero-proxy methods. --- **Thirdly**, the computational cost aspect of POMONAG is weak. The section "Generation and Training Time" should be better presented. The method requires a diffusion generation phase which takes extra time. That itself is a disadvantage. Also timewise, POMONAG cannot claim superiority as recent methods mentioned earlier are faster. * Present a detailed table comparing computational costs (including generation and training time) of POMONAG with other methods, including these zero proxy methods mentio