Second language learners reach expert levels in relative cue weighting only gradually. On the basis of ensemble machine learning models fit to naturalistic written productions of German advanced learners of English and expert writers, we set out to reverse engineer differences in the weighting of multiple cues in a clause linearization problem. We found that, while German advanced learners succeeded in identifying important cues, their assignment of cue importance differed from that of the expert control group. Even at advanced levels, learners are found to rely on a smaller set of perceptually salient cues than native speakers do, focusing on cues that exhibit relatively high cue availability and relatively low cue reliability. Our findings suggest that the principles of the Unified Model of first and second language acquisition, which have been extensively supported for comprehension also underlie the written production of advanced second language learners.