An enlargement of the candidate vector set of QR-least reliable layer (QR-LRL) based MIMO detector for efficient soft output generation is proposed. Previous work (Bahng et al. in IEICE Trans Commun, E89–B(10):2956–2960, 2008) shows that the QR-LRL based MIMO detector approaches hard decision output ML performance, but does not match soft output ML performance due to empty candidate vector set problem. Performance degradation is more severe when modulation order is low. Some of the previous methods have provided solutions to empty vector set (EVS) problem (Kawai et al. in IEICE Trans Commun, E88–B(1):47–57, 2005; Bahng et al. in IEICE Trans Commun, E89–B(10):2956–2960, 2008; Kim et al. in IEICE Trans Commun, E92–B(11):3512–3515, 2009), but are not efficient in terms of performance or computation complexity. In this paper, we enlarge the candidate vector set of QR-LRL detector by applying every constellation point at each layer. The proposed detector thus effectively solves the EVS problem and achieves soft ML performance while keeping the computation complexity low, especially at low modulation order. © 2015, Springer Science+Business Media New York.