Sphere decoding has been used for maximum likelihood (ML) detection of multiple-input-multiple-output (MIMO) communication systems. In order to combine the sphere decoder with the outer channel code decoder, researchers have proposed to include a candidate list in the sphere decoder to provide soft information to the channel code decoder. This algorithm is called list sphere decoder (LSD). With the LSD, due to the introduction of the candidate list, the decoding speed is around 10 times lower than that of the conventional sphere decoder (SD) if the LSD uses the largest Euclidean distance (ED) in the list for the radius reduction. At the same time, if the candidate list is divided into multiple sublists to reduce the complexity of the list updating circuit, the average number of cycles required becomes even larger. Meanwhile, in real applications, the maximal number of cycles in the sphere decoder is usually bounded to guarantee a certain data throughput. This bound will also affect the performance of the LSD. In this paper, we propose a radius reduction method which uses either the mean value or the median value of the Euclidean distances (ED) of the candidates in the list. This algorithm leads to a faster decoding speed with little or no performance loss. If the idea of using sublists is applied, this radius reduction method reduces the amount of increase in the average number of cycles. Furthermore, if a bound is set on the maximal number of cycles, this method can provide better performance than the original LSD. We also propose a simple performance evaluation method for comparing different radius reduction methods without involving a real soft-input-soft-output channel code decoder.