Mountains in arid regions are undergoing rapid climate warming along distinct altitudinal gradients. However, the impact of this warming on soil carbon cycling remains unclear. In this study, we combined extensive field sampling and machine learning to reconstruct the spatiotemporal dynamics (1993-2023) of soil organic carbon stocks (SOCS) across altitudinal vegetation belts in the Tianshan Mountains of central Asia. Our results showed that SOCS decreased in low-elevation vegetation belts, that is, desert and montane steppe, and increased in highelevation vegetation belts in coniferous forests and alpine meadows. SOCS significantly declined in the 0-10 cm soil layer of the montane steppe, whereas it increased in the 10-30 cm soil layer of the coniferous forests. Warming was identified as the dominant driver of the interannual SOCS trends across both vegetation belts. These findings have shown contrasting soil organic carbon (SOC) responses along elevational gradients and highlighted the role of high-elevation vegetation, particularly coniferous forests, as stable carbon sinks. Our study has provided the first long-term, spatially explicit SOCS estimates for this arid mountainous region, offering insights into soil carbon-climate feedback under ongoing warming.