The sphere decoder (SD) has been utilized for maximum likelihood (ML) detection in MIMO systems. In order to improve system performance, the SD is usually combined with the error correction codes where soft decoding is utilized. The list sphere decoder (LSD) was introduced to generate a candidate list, which can provide soft information to the outer decoder. Unlike the conventional sphere decoder, the LSD has the candidate list updating and new radius generation units, which causes extra complexity and latency. In this paper, we present a novel radius updating architecture, which can obtain a new radius much faster than the conventional method. Furthermore, we propose an efficient candidate list updating scheme, which can significantly save the complexity (without affecting the decoding speed) of updating the candidate list used to compute soft information.