1. Great efforts have been made to control soil erosion by restoring plant communities in degraded ecosystems world-wide. However, soil erosion has not been substantially reduced mainly because current restoration strategies lead to large areas of mono-specific vegetation, which are inefficient in reducing soil erosion because of their simple canopy and root structure. Therefore, an advanced understanding of how community functional composition affects soil erosion processes, as well as an improved restoration scheme to reduce soil erosion, is urgently needed. 2. We investigated the effect of community functional composition on soil erosion in restored semi-arid grasslands on the Loess Plateau of China. Community functional composition of 16 restored grasslands was quantified by community-weighted mean (CWM) and functional diversity (FD) trait values, which were calculated from nine plant functional traits of thirteen locally dominant plant species. Species richness and evenness were also measured. Soil erosion rates were measured using standard erosion plots. The multimodel inference approach was used to estimate the direction and the relative importance of these biodiversity indices in reducing soil erosion. 3. A robust and strong negative effect of functional divergence (FDiv) on soil erosion was found. The prevalence of particular trait combinations can also decrease soil erosion. The greatest control over soil erosion was exerted when the community mean root diameter was small and the root tensile strength was great. 4. Synthesis and applications: These findings imply that community functional diversity plays an important role in reducing soil erosion in semi-arid restored grasslands. This means that current restoration strategies can be greatly improved by incorporating community functional diversity into restoration design. We propose a trait-based restoration framework for reducing soil erosion, termed ‘SSM’ (Screening–Simulating–Maintaining). SSM aims to translate the target of community functional diversity into community assemblages that can be manipulated by practitioners. Based on this framework, a comprehensive procedure, highlighting functional diversity as the primary concern in determining optimal community assemblages, was developed to meet the pressing need for more effective restoration strategies to reduce soil erosion.