Abstract: Climate model simulations are generally found to exhibit poorer performance in simulating compound extreme events than in simulating single-variable extremes, which leads to greater uncertainties in future projections. In this study, univariate quantile delta mapping (QDM) and a multivariate bias correction method (MBCn) are applied to adjust biases in surface air temperature, precipitation, and the compound hot–dry index (PI) across China. The bias correction is performed using simulations from the regional climate model RegCM4.4, driven by three global climate models participating in CMIP6. The performance of the two methods is compared and evaluated. After the superiority of MBCn in correcting compound variables is validated, the projected changes in compound hot–dry events over China are analyzed under the SSP245 scenario for the mid-term(2021–2060) and long-term (2061–2100) future periods. Both QDM and MBCn showed excellent performance in correcting the univariate biases of temperature and precipitation, with correlation coefficients exceeding 0.99 and 0.80, respectively, and mean biases constrained within ±0.5°C and ±1mm/d. However, limited performance is shown by QDM in correcting the PI index, whereas the joint distribution characteristics among variables are effectively preserved by MBCn, increasing the spatial correlation coefficient to 0.71–0.97 and reducing the mean bias and root-mean-square error, thereby demonstrating a clear advantage in simulating compound events. Under the SSP245 scenario, the extremity of compound hot–dry events across China is projected to intensify, with the most pronounced increase in risk occurring in winter. A shift from a nationwide expansion in the mid-term to a concentration in northern China in the long-term is exhibited in the spatial pattern of risk evolution, with Central China and East China emerging as high-risk hotspots where risks from multiple seasons overlap.