A structural approach to multi-stage detection is proposed where the joint information between interfering users is utilized to give better soft decisions on the user bits in each stage or iteration. The key idea is to use the structure of the multi-user signal constellation to achieve higher performance in terms of the bitor-rate, particularly at high SNR, and reduce the possibility of error propagation or limit cycles. A maximal asymptotic efficiency detector is employed as the first stage to generate better initial estimates of the user bits, followed by a local maximum likelihood kernel for subsequent stages. The proposed detector is particularly attractive for heavily correlated or over-loaded multi-user systems. The proposed multi-stage detector also includes the possibility of known memoryless non-linearity in the system, e.g., the non-linearity introduced due to saturation by the downlink RF amplifier in satellite communication systems.