This paper presents a novel probabilistic spherical detection (P-SD) method which applies the probabilistic-search algorithm to conventional depth-first SD (DF-SD). By confining the tree search into candidates which can be selected in an adaptive manner, a large number of promising candidates can be evaluated before termination. Consequently, the proposed P-SD improves the error performance of DF-SD with early termination, while retaining the hardware efficiency. An efficient VLSI architecture is proposed for implementation of the P-SD algorithm, and the results of the synthesized architecture are presented. The main advantage of P-SD is that it can fully exploit the state-of-the-art architectures of DF-SD, since it can be implemented by simply adding two functional blocks to conventional DF-SD. By analyzing the performance-complexity tradeoffs, it is concluded that our proposed P-SD is advantageous over conventional DF-SD and K-best algorithm, when the maximum-likelihood error performance is desired.