SasWOT is a novel training-free framework for semantic segmentation architecture search (SAS) that leverages an auto-discovery proxy to improve search efficiency. The framework addresses the computational costs and practical limitations of traditional training-based SAS methods by exploring a customized proxy search space. SasWOT combines multiple statistics (weights, gradients) as inputs and advanced mathematical operations to enhance the predictive capabilities of the proxies. An evolutionary algorithm is employed to evolve the best-performing proxies, which are then used to search for optimal segmenters without candidate training. Extensive experiments on the Cityscapes and CamVid datasets demonstrate that SasWOT achieves superior trade-offs between accuracy and speed, with a performance of 71.3% mIoU and a speed of 162 FPS on Cityscapes. SasWOT not only enables automated proxy optimization but also significantly reduces the search time, making it a practical solution for real-time applications.SasWOT is a novel training-free framework for semantic segmentation architecture search (SAS) that leverages an auto-discovery proxy to improve search efficiency. The framework addresses the computational costs and practical limitations of traditional training-based SAS methods by exploring a customized proxy search space. SasWOT combines multiple statistics (weights, gradients) as inputs and advanced mathematical operations to enhance the predictive capabilities of the proxies. An evolutionary algorithm is employed to evolve the best-performing proxies, which are then used to search for optimal segmenters without candidate training. Extensive experiments on the Cityscapes and CamVid datasets demonstrate that SasWOT achieves superior trade-offs between accuracy and speed, with a performance of 71.3% mIoU and a speed of 162 FPS on Cityscapes. SasWOT not only enables automated proxy optimization but also significantly reduces the search time, making it a practical solution for real-time applications.