Orthogonal Least Squares Regression (OLSR) selects each regressor by repeated weighted boosting search (RWBS). This kind of OLSR is known to be capable of producing a much sparser model than many other kernel methods. With the aid of tree structure search, this paper is to construct an even sparser regression model in the framework of OLSR with RWBS. When RWBS being used to solve the optimization at each regression stage, OLSR is extended by keeping the k ( k >1)excellent regressors, which minimize the modeling MSE, rather than only choose the best one at each iteration. In this way, the next regressor will be searched in k subspaces instead of in only one subspace as the conventional method. Furthermore we propose a subtree search to decrease experimental time complexity, by specifying the total number of children in every tree depth. The new schemes are shown to outperform the traditional method in the applications, such as component detection, sparse representation for ECG signal and 2-d time series modeling. Besides, experimental results also indicate that subtree based algorithm is with much lower time complexity than tree based one.