Soil erosion, as a global ecological and environmental issue, severely threatens soil productivity and ecosystem stability. Investigating its spatiotemporal dynamics and driving mechanisms holds significant scientific value for the sustainable development of ecologically fragile regions. To address the limitations of traditional assessment methods in capturing dynamic response characteristics and evaluating ecosystem resilience, this study proposed a soil erosion vulnerability assessment framework based on the "Frequency-Magnitude-Resilience" paradigm. Utilizing multi-source spatiotemporal data from 1990 to 2022 and integrating the optimal parameter-based geographical detector model, this study systematically analyzed the evolution of soil erosion and its driving mechanisms in Yunnan Province, a typical ecologically fragile region. The results indicated that: (1) Mild erosion dominated in Yunnan Province, accounting for over 76.80%, with a positive trend of increasing low-intensity erosion and decreasing high-intensity erosion. (2) The spatial differentiation of soil erosion vulnerability was pronounced, with stable and relatively stable zones dominating (58.21%), while certain localized areas exhibited high vulnerability (10.84%). (3) Slope had the strongest explanatory power on the spatial differentiation of vulnerability (q = 0.2240). (4) The driving mechanisms exhibited a gradient response pattern: soil erosion in stable and relatively stable zones was primarily regulated by temperature and vegetation cover; in moderately vulnerable zones, it was jointly driven by vegetation cover and soil organic carbon content; in relatively and extremely vulnerable zones, precipitation and slope were the dominant factors, with soil properties and vegetation cover playing an increasingly significant role in extremely vulnerable areas. The "Frequency–Magnitude–Resilience" vulnerability identification method proposed in this study provided a novel theoretical perspective and technical tool for soil erosion assessment.