Researchers have developed a new method for discovering ensemble zero-cost proxies in neural architecture search (NAS). These proxies enable NAS to rank candidate networks based on statistics computed at initialization, without the need for repeated training. The proposed method, Bi-EZP, uses a large language model to generate executable aggregation programs over four base proxies, and then optimizes their parameters using an evolutionary strategy. Experiments show that this approach can improve ranking performance across different search spaces. --- Why it matters: This matters because it provides a more efficient way to discover ensemble zero-cost proxies in NAS, which can speed up the architecture search process and improve the quality of discovered networks. Source: https://arxiv.org/abs/2608.21927