The Revised Universal Soil Loss Equation, Version 2 (RUSLE2) is widely used for regional soil erosion estimation. However, its performance on croplands in the Mollisol region of Northeast China remains insufficiently quantified, particularly for areas characterized by long and gentle (> 100 m and < 10 degrees) and with different slope shapes (convex, straight, concave). This study integrated (1) Cs-3(7) tracing technique to identify systematic underestimation errors in RUSLE2 predictions, primarily caused by oversimplified linear assumptions in conventional slope length factor (lambda) calculations. To address this limitation, this study employed Random Forest Regression (RFR) to model non-linear lambda- relationships, where represents the slope shape weight factor. The results revealed that the distribution of predicted and measured erosion-deposition rates across different slope shapes exhibited a clear alternating pattern of strong and weak values. Furthermore, the fluctuations in the predicted values were aligned with alterations in the slope gradient. Both predicted and measured mean erosion rates on different slopes were in the order that convex > straight > concave. The optimized model significantly improved predictive performance, with slope length (lambda) and slope shape factor () as the key factors. The model demonstrated strong adaptability, with the highest predictive accuracy achieved for concave (R-2 = 0.99*), followed by straight (R-2 = 0.88*) and convex (R-2 = 0.82*) slopes. This study provides theoretical support for improving soil erosion prediction models and a scientific basis for optimizing soil and water conservation strategies.