DartsNeXt: Bridging the Search-Evaluation Gap in Differentiable NAS With Router-Based Selection and Sequential Architecture

Neural Architecture Search (NAS) has emerged as a pivotal approach for automating the design of high-performance neural networks, with Differentiable Architecture Search (DARTS) widely being recognized for its efficiency. However, DARTS suffers from critical limitations, including a soft–hard mismatch and a depth mismatch between the search and evaluation phases, leading to instability and performance collapse. To address these issues, we propose DartsNeXt, a sequential differentiable NAS framework that aligns search and evaluation across both the operator and macro levels. At the operator level, we introduce a novel router-based selection method that replaces magnitude-based selection with input-conditioned sparse routing, ensuring that the forward graph during search is consistent with the final discrete architecture while reducing bias toward skip connections. At the macro level, a stage-wise backbone with large-kernel depthwise convolutions minimizes memory usage and mitigates the depth discrepancy, enabling search to be performed directly at the evaluation depth. Extensive experiments demonstrate that DartsNeXt achieves stable and robust performance on NAS-Bench-201 and the DARTS search space. Furthermore, DartsNeXt provides a fully automated and compute-efficient pipeline that discovers competitive token mixers without manual design, delivering near-expert performance on ImageNet-1 K classification and COCO detection tasks under realistic compute budgets.