When the fitness landscape exhibits a multi-modal property, clustering plays a key role in the evolutionary computation, because clusters explicitly or implicitly denote optima present. Correct clusters result in effective and efficient evolution. In this paper, a novel clustering strategy, called Recursive Middling (RM), is proposed. With acceptable overhead, RM effectively overcomes pitfalls of other popular clustering techniques, i. e. those based on Euclidean distance or Hill-Valley function [1]. RM also dramatically enhances the performance of the selected evolutionary algorithm – Dynamic Niche Clustering (DNC) [2], by forming clusters centered around potential optima quickly and stably. The success rate and the number of optima found are both increased dramatically, compared to the original version of DNC.