The Revised Universal Soil Loss Equation (RUSLE), a globally recognized empirical model for soil erosion assessment, has seen extensive efforts to calibrate its erosion factors (R, K, LS, C, P) across diverse regions and scales. However, systematic evaluations of how factor configurations influence model performance, particularly in revegetated ecosystems remain scarce. To address this gap, this study analyzed 8 years of monitoring data (2016 similar to 2023) from 10 runoff plots in a revegetated watershed on China's Loess Plateau, where vegetation restoration has dramatically altered erosion dynamics. We rigorously tested 144 RUSLE configurations at annual/multi-year scales and 36 configurations at rainfall event scales, evaluating performance through Nash-Sutcliffe Efficiency (NSE) and Mean Absolute Percentage Error (MAPE). Results of the 144 factor configurations revealed systematic bias of soil loss rates across revegetated plots, with NSE values spanning -41.48 to -0.11 and MAPE ranging from 143.22% to 1,540.23%. Performance varied markedly across temporal scales: annual/multi-year simulations (NSE: -4.17 similar to-0.52; MAPE: 91.96 similar to 554.90%) outperformed event-scale predictions (NSE: -8.04 similar to-0.63; MAPE: 99.69 similar to 709.25%), aligning with RUSLE's original design for long-term averages. Rainfall intensity further modulated accuracy, as heavy rainfall events (NSE: -7.61 similar to-2.39; MAPE: 316.92 similar to 938.50%) exhibited larger errors than non-heavy events (NSE: -1.27 similar to-0.56; MAPE: 219.17 similar to 515.72%), highlighting the model's inability to resolve intensity-dependent thresholds. Optimized configurations, such as event-scale R factor refinements (R-W) coupled with vegetation-adaptive C factors (C-L) reduced uncertainties effectively, demonstrating the value of localized factor calibration. However, persistent biases highlighted inherent limitations in RUSLE's linear empirical framework, which oversimplifies nonlinear interactions between vegetation, soil, and rainfall. These findings emphasize the need for context-driven factor selection to enhance RUSLE's utility in ecological restoration regions, while advocating for future integration with process-based models to address mechanistic gaps in dynamic, revegetated landscapes.