Multiple unordered treatments with a binary instrument for each treatment are common in policy evaluation. This multiple treatment setting allows for different types of changes in treatment status that are non-compliant with the activated instrument. Therefore, instrumental variable (IV) methods have to rely on strong assumptions on the subjects' behavior to identify local average treatment effects (LATEs). This paper introduces a new IV strategy that identifies an interpretable weighted average of LATEs under relaxed assumptions, in the presence of clusters with similar treatments. The clustered LATEs allow for shifts across treatment clusters that are consistent with preference updating, but render IV estimation of individual LATEs biased. The clustered LATEs are estimated by standard IV methods, and we provide an algorithm that estimates the treatment clusters. We empirically analyze the effect of fields of study on academic student progress, and find violations of the LATE assumptions in line with preference updating, clusters with similar fields, treatment effect heterogeneity across students, and significant differences in student progress due to fields of study.