The computational complexity of a sphere decoder (SD) is conventionally reduced by decoding the order scheme which sorts candidate symbols in the ascending order of the Euclidean distance from the soft-output signal points of a zero-forcing (ZF) receiver. However, since the soft-output signal point of ZF may not be a reliable sorting reference, we propose two types of sorting schemes to allow faster decoding. The first is to use the newly found lattice points in the previous search round instead of the ZF solution (Type I). Since these lattice points are closer to the received signal than the ZF solution, they can serve as a more reliable sorting reference for finding the maximum likelihood (ML) solution. The second sorting scheme is to sort candidate symbols in descending order according to the number of candidate symbols in the following layer, which are called child symbols (Type II). These two proposed sorting schemes can be combined with layer sorting for more complexity reduction. Through simulation, the Type I and Type II sorting schemes were found to provide 12% and 20% complexity reduction respectively over conventional sorting schemes. When they are combined with layer sorting, Type I and Type II provide an additional 10-15% complexity reduction while maintaining detection performance.